If `eager_loading` is true, `apply_join_dependency` force applies
LIMIT/OFFSET before JOINs by `limited_ids_for` to keep parent records
count. But for aggregation queries, LIMIT/OFFSET should be applied after
aggregations the same as SQL semantics.

And also, we could not replace SELECT list by `limited_ids_for` when a
query has a GROUP BY clause. It had never been worked since it will
causes generating invalid SQL for MySQL, PostgreSQL, and probably most
backends.

```
% ARCONN=postgresql be ruby -w -Itest test/cases/calculations_test.rb -n test_group_by_with_limit
Using postgresql
Run options: -n test_group_by_with_limit --seed 20925

# Running:

E

Error:
CalculationsTest#test_group_by_with_limit:
ActiveRecord::StatementInvalid: PG::GroupingError: ERROR:  column "posts.id" must appear in the GROUP BY clause or be used in an aggregate function
LINE 1: SELECT  DISTINCT "posts"."id", "posts"."type" AS alias_0 FRO...                         ^
: SELECT  DISTINCT "posts"."id", "posts"."type" AS alias_0 FROM "posts" LEFT OUTER JOIN "comments" ON "comments"."post_id" = "posts"."id" GROUP BY "posts"."type" ORDER BY "posts"."type" ASC LIMIT $1
```

Fixes #8103.
Closes #27249.

In `_create_record`, explicit `transaction` block requires rollback
handling manually when `insert_record` is failed.

We need to handle it in `_create_record`, not in `insert_record`, since
our test cases expect a record added to target and returned even if
`insert_record` is failed,

Closes #31488.

Follow up of #32952.

Context: https://github.com/rails/rails/commit/43ef00e5d7a55ad79bc840276d33cb70f1f5dde5#commitcomment-29256140

Rails 5.2 does not alias child joins, causing an error about duplicated table/fields:

Example:

Using some code like:
`Post.joins(:author, :categorizations).merge(Author.select(:id)).merge(Categorization.joins(:author))`

*Before this fix:*
`
SELECT ... FROM "posts" INNER JOIN "authors" ON ... INNER JOIN "authors" ON ...
`

*After this fix:*
`
SELECT ... FROM "posts" INNER JOIN "authors" ON ... INNER JOIN "authors" "authors_categorizations" ON ...
`

Before 5.2, Rails aliased the joins, but wrongfully transformed them into a LEFT OUTER JOIN.
This fix will keep them as INNER JOINS, but make sure child joins are aliased, to avoid errors.

Should be done before `before_add` callbacks.

Since a record is already persisted in `after_create_commit`, so `save`
should invoke only `after_update_commit`.

This bug is caused by depending on `@_start_transaction_state` for
rollback to consider whether it was `new_record` before being committed.

If after commit callbacks caused another commit, the state before
last commit is no longer `new_record`.

Fixes #32831.
Closes #18367.
Closes #31106.

Fixes #32806.

- MariaDB 10.3.7 is the first GA release
https://mariadb.com/kb/en/library/mariadb-1037-release-notes/

- MariaDB 10.3 translates `LENGTH()` to `OCTET_LENGTH()` function

https://mariadb.com/kb/en/library/sql_modeoracle-from-mariadb-103/
> MariaDB translates LENGTH() to OCTET_LENGTH()

- MySQL does NOT translate `LENGTH()` to `OCTET_LENGTH()`
However, it translates `OCTET_LENGTH()` to `LENGTH()`

Here are generated schema dumps of this test to show the differences
between MySQL and MariaDB:

* MySQL 8.0 (Server version: 8.0.11 MySQL Community Server - GPL)
```ruby
  create_table \"virtual_columns\", options: \"ENGINE=InnoDB DEFAULT CHARSET=utf8 COLLATE=utf8_unicode_ci\", force: :cascade do |t|
    t.string \"name\"
    t.virtual \"upper_name\", type: :string, as: \"upper(`name`)\"
    t.virtual \"name_length\", type: :integer, as: \"length(`name`)\", stored: true
    t.virtual \"name_octet_length\", type: :integer, as: \"length(`name`)\", stored: true
  end
```

* Maria DB 10.3 (Server version: 10.3.7-MariaDB-1:10.3.7+maria~bionic-log mariadb.org binary distribution)

```ruby
  create_table \"virtual_columns\", options: \"ENGINE=InnoDB DEFAULT CHARSET=utf8 COLLATE=utf8_unicode_ci\", force: :cascade do |t|
    t.string \"name\"
    t.virtual \"upper_name\", type: :string, as: \"ucase(`name`)\"
    t.virtual \"name_length\", type: :integer, as: \"octet_length(`name`)\", stored: true
    t.virtual \"name_octet_length\", type: :integer, as: \"octet_length(`name`)\", stored: true
  end
```

`QueryAttribute#value_for_database` calls only `type.serialize`, and
`Boolean#serialize` is a no-op unlike other attribute types.

It caused the issue #32624. Whether or not `serialize` will invoke
`cast` is undefined in our test cases, but it actually does not work
properly unless it does so for now.

Fixes #32624.

- Fixes #32940

Since `parse_raw_value_as_a_number` may not always parse raw value from
database as a number without type casting (e.g. "$150.55" as money
format).

Fixes #32531.

Since #31405, using `#increment!` with touch option instead of `#touch`
to touch belongs_to association if counter cache is enabled. It caused
the regression since `#increment!` won't invoke after_touch callbacks
even if touch option is given.
To fix the regression, make `#increment!` invokes after_touch callbacks
if touch option is given.

Fixes #31559.
Fixes #32408.

On belongs_to with `touch: true` association, unassigned object is
caused touching, but assigned object is not touched.

And also, if primary key is customized, it will touch against the wrong
target looked up by the customized key as primary key.

This change ensures correctly touching consistently between assigning
and unassigning.

If belongs_to primary key is customized, the callback will update
counters against the wrong target looked up by the customized key as
primary key.
We need to convert the customized key into an object that can be
referred to as primary key.

Since #31575, `BelongsToAssociation#target=` replaces owner record's
foreign key to fix an inverse association bug.

But the method is not only used for inverse association but also used
for eager loading/preloading, it caused some public behavior changes
(#32338, #32375).

To avoid any side-effect in loading associations, I reverted the
overriding `#target=`, then introduced `#inversed_from` to replace
foreign key in `set_inverse_instance`.

Closes #32375.

Since #26074, introduced force equality checking to build a predicate
consistently for both `find` and `create` (fixes #27313).

But the assumption that only array/range attribute have subtype was
wrong. We need to make force equality checking more strictly not to
allow serialized attribute.

Fixes #32761.

* Rollback parent transaction when children fails to update

Rails supports autosave associations on the owner of a `has_many`
relationship. In certain situation, if the children of the association
fail to save, the parent is not rolled back.

```ruby
class Employee < ActiveRecord::Base
end

class Company < ActiveRecord::Base
  has_many(:employees)
end

company = Company.new
employee = company.employees.new
company.save
```

In the previous example, if the Employee failed to save, the Company
will not be rolled back. It will remain in the database with no
associated Employee.

I expect the `company.save` call to be atomic, and either create all or
none of the records.

The persistance of the Company already starts a transaction that nests
it's children. However, it didn't track the success or failure of it's
children in this very situation, and the outermost transaction is not
rolled back.

This PR makes the change to track the success of the child insertion and
rollback the parent if any of the children fail.

* Change the test to reflect what we expect

Once #32862 is merged, rolling back a record will rollback it's state to match
the state before the database changes were applied

* Use only the public API to express the tests

* Refactor to avoid reassigning saved for nested reflections

[Guillaume Malette + Rafael Mendonça França]

Unlike other databases, changing SQLite3 table definitions need to create a temporary table.
While changing table operations, the original table needs dropped which caused
`SQLite3::ConstraintException: FOREIGN KEY constraint failed` if the table is referenced by foreign keys.
This pull request disables foreign keys by `disable_referential_integrity`.

Also `disable_referential_integrity` method needs to execute `defer_foreign_keys = ON`
to defer re-enabling foreign keys until the transaction is committed.

https://www.sqlite.org/pragma.html#pragma_defer_foreign_keys

Fixes #31988

- This `defer_foreign_keys = ON` has been supported since SQLite 3.8.0
https://www.sqlite.org/releaselog/3_8_0.html and Rails 6 requires SQLite 3.8 #32923 now

- <Models>.reset_column_information added to address `ActiveModel::UnknownAttributeError`

```
Error:
ActiveRecord::Migration::ForeignKeyChangeColumnTest#test_change_column_of_parent_table:
ActiveModel::UnknownAttributeError: unknown attribute 'name' for ActiveRecord::Migration::ForeignKeyChangeColumnTest::Post.
```

To prevent redundant `to_s` like https://github.com/rails/rails/pull/32923#discussion_r189460008
automatically in the future.

These OS versions have SQLite 3.8 or higher by default.

- macOS 10.10 (Yosemite) or higher
- Ubuntu 14.04 LTS or higher

Raising the minimum version of SQLite 3.8 introduces these changes:

- All of bundled adapters support `supports_multi_insert?`
- SQLite 3.8 always satisifies `supports_foreign_keys_in_create?` and `supports_partial_index?`
- sqlite adapter can support `alter_table` method for foreign key referenced tables by #32865
- Deprecated `supports_multi_insert?` method

After a real (non-savepoint) transaction has committed or rolled back,
the original persistence-related state for all records modified in that
transaction is discarded or restored, respectively.

When the model has transactional callbacks, this happens synchronously
in the `committed!` or `rolled_back!` methods; otherwise, it happens
lazily the next time the record's persistence-related state is accessed.

The synchronous code path always finalizes the state of the record, but
the lazy code path only pops one "level" from the transaction counter,
assuming it will always reach zero immediately after a real transaction.
As the test cases included here demonstrate, that isn't always the case.

By using the same logic as the synchronous code path, we ensure that the
record's state is always updated after a real transaction has finished.

Before it was coercing an invalid string into "2000-01-01 00:00:00".

After #449 was merged math can be done on these
nodes, adding a test file to unit test all the
math operators.

Follow up of #32605.

`becomes` creates new object and copies attributes from the receiver. If
new object has mutation tracker which is created in `after_initialize`,
it should be cleared since it is for discarded attributes.

But if the receiver doesn't have mutation tracker yet, it will not be
cleared properly.

It should be cleared regardless of whether the receiver has mutation
tracker or not.

Fixes #32867.

If an `ActiveRecord::Rollback` error was raised by a persistence method
(e.g. in an `after_save` callback), this logic would potentially discard
the original state of the record from before the transaction, preventing
it from being restored later when the transaction was rolled back.

https://travis-ci.org/rails/rails/jobs/375326992#L1160-L1166

Commit callbacks are intentionally disabled when errors occur when calling the callback chain in order to reset the internal record state. However, the implicit order of operations on the logic for checking if callbacks are disabled is wrong. The result is that callbacks can be unexpectedly when errors occur in transactions.

It had been added at https://github.com/rails/arel/commit/05b5bb12270b32e094c1c879273e0978dabe5b3b
and removed at https://github.com/rails/arel/commit/db1bb4e9a728a437d16f8bdb48c3b772c3e4edb0

`require 'rubygems'` is already required in Ruby 1.9 or later.

Ensure that do not accidentally remove an index of different
definitions.

If a 'has one' object is created from a new record, an ActiveRecord::RecordNotSaved error is raised but this behavior was also applied to the reverse scenario.

It will cause "undefined method `test_order' for ActiveSupport:Module
(NoMethodError)".

https://travis-ci.org/rails/rails/jobs/373472604#L1208