when model is added to collection association

Follow up of #31390.

Otherwise random CI failure will be caused.

https://travis-ci.org/rails/rails/jobs/265848046#L777

It is incorrect to treat `options[:primary_key]` as
`association_primary_key` if `has_many` associations because the
`:primary_key` means the column on the owner record, not on the
association record. It will break `ids_reader` and `ids_writer`.

```ruby
  people(:david).essay_ids
  # => ActiveRecord::StatementInvalid: Mysql2::Error: Unknown column 'essays.first_name' in 'field list': SELECT `essays`.first_name FROM `essays` WHERE `essays`.`writer_id` = 'David'
```

Fixes #14439.

Since 07e5301e, `Relation#first` will order by primary key if no order is
defined.

This reverts commit 3420a145, reversing
changes made to afb66a5a.

Fixes #29627.

Unscoping `default_scope` in associations has already supported (#17360
for preloading, c9cf8b8d for eager loading).

Fixes #20679.
Closes #16531.

And use `assert_same` instead of `assert_equal` and tiny fix assert
message s/#reload/#reset/.

Follow up of #29511.

The `@offsets` cache is used by FinderMethods to cache records found by
find_nth. This cache is cleared in AR::Relation#reset, but not in
CollectionProxy#reset or CollectionProxy#reload.

Because of this, since #29098, calling #first/#find_nth/etc after
calling #reload or #reset on an association could return a stale record.
This is an issue both when the full association target is loaded and
when the item is loaded in #find_nth.

This commit solves the problem by clearing the `@offsets` cache in
CollectionProxy#reset and CollectionProxy#reload.

This fixes the following issues.

* `association_scope` doesn't include `default_scope`. Should use `scope` instead.
* We can't use `method_missing` for customizing existing method.
* We can't use `relation_delegate_class` for sharing extensions. Should extend per association.

This change reverted in eac6f369 but it is needed for data integrity.
See #25328.

- Tried specifying `id` for the `readers` records but it is
  interconnected with so many tests that many random tests started
  failing.
- So switched to the approach of deleting all readers in the
  `create_resets_cached_counters` test.

Revert "Merge pull request #27636 from mtsmfm/disable-referential-integrity-without-superuser-privilege-take-2"

This reverts commit c1faca63, reversing
changes made to 8c658a0e.

See https://github.com/rails/rails/pull/27636#issuecomment-297534129

Cannot call private methods in `@klass` against `CollectionProxy`
(inherites `Relation`) because using `public_send` in `method_missing`.

Follow up of 03d3f036.

Some of `respond_to?` were replaced to `respond_to_missing?` in 03d3f036.
But the visibility is still public. It should be private.

Prevent double firing the before save callback of new object when the parent association saved in the callback

Related #18155, #26661, 268a5bb0, #27434, #27442, and #28599.

Originally #18155 was introduced for preventing double insertion caused
by the after save callback. But it was caused the before save issue
(#26661). 268a5bb0 fixed #26661, but it was caused the performance
regression (#27434). #27442 added new record to `target` before calling
callbacks for fixing #27434. But it was caused double firing before save
callback (#28599). We cannot add new object to `target` before saving
the object.

This is improving #18155 to only track callbacks after `save`.

Fixes #28599.

Add regression test for setting inverse instances on normal & polymorphic relationships when building objects on new records

Re-create https://github.com/rails/rails/pull/21233

eeac6151 was reverted (127509c0) because it breaks tests.

----------------

ref: 72c15572

- We must use `authors` fixture with `author_addresses` because of its foreign key constraint.
- Tests require PostgreSQL >= 9.4.2 because it had a bug about `ALTER CONSTRAINTS` and fixed in 9.4.2.

Since b644964b `ActiveRecord::Relation` includes `Enumerable` so
delegating `collect`, `all?`, and `include?` are also unneeded.
`collect` without block returns `Enumerable` without preloading by that.
We should use `load` rather than `collect` for force loading.

The `select` in `QueryMethods` is also an enumerable method.
Enumerable methods with block should delegate to `records` on
`CollectionProxy`, not `scope`.

Fixes #28348.

I incorrectly changed behavior of `dup`. Reading the original issue I
thought that `dup` should retain the original contents of the record
and it's associations but it is in fact supposed to be a copy as if a
record had been reinitialized.

This reverts commit ca8c21df.

In Rails 3.2 dupping a `CollectionProxy` would dup it's `load_target` as
well. That functionality has been broken since the release of Rails 4.0.
I hit this in an application upgrade and wondered why duplicating a
CollectionProxy and assigning it to a variable stopped working.

When calling `dup` on a `CollectionProxy` only the owner (ex.
topic) was getting duplicated and the `load_target` would remain in tact
with it's original object ID. Dupping the `load_target` is useful for performing
a logging operation after records have been destroyed in a method.

For example:

```
def transfer_operation
  saved_replies = topic.replies

  topic.replies.clear

  saved_replies.each do |reply|
    user.update_replies_count!
  end
end
```

This change adds a `initialize_dup` method that performs a `deep_dup` on
the `@associatiation` so that the `load_target` is dupped as well.

Fixes #17117

Problem: Calling `count` on an association can cause invalid SQL queries
to be created where the `SELECT COUNT(a, b, c)` function receives
multiple  columns. This will cause a `StatementInvalid` exception later
on.

Solution: Use `count(:all)`, which generates a `SELECT COUNT(*)...`
query independently of the association.

This also includes a test case that, before the fix, broke.

Revert "Merge pull request #21233 from mtsmfm/disable-referential-integrity-without-superuser-privileges"

This reverts commit eeac6151, reversing
changes made to 5c40239d.

Reason: Broke the isolated tests.
https://travis-ci.org/rails/rails/builds/188721346

Because merging the association scope was removed.

`append_record` was added at 15ddd517 for not double adding the record.
But adding `append_record` (checking `@target.include?(record)`) caused
performance regression #27434. Instead of checking not double adding the
record, add a record to target before any callbacks loads the record.

Fixes #27434.