Fix `modules_for_helpers` to will raise `MissingHelperError`

Deprecate marshalling load from legacy attributes format

Motivation is twofold:

  * We are gradually removing `require_dependency` from the framework.

  * Let `helper` work if `config.add_autoload_paths_to_load_path` is
    disabled.
Co-authored-by: NJean Boussier <jean.boussier@gmail.com>

Let's bump Zeitwerk in preparation for Rails 6.1.

With Zeitwerk 2.3, applications can enable reloading and eager loading
at the same time. As of this writing, Rails does not implement that
logic, but if we do we know the dependency is in place to support it.

Zeitwerk minor releases are backwards compatible, should be a seamlessly
upgrade.

Since #31827, marshalling attributes hash format is changed to improve
performance because materializing lazy attribute hash is too expensive.

In that time, we had kept an ability to load from legacy attributes
format, since that performance improvement is backported to 5-1-stable
and 5-0-stable.

Now all supported versions will dump attributes as new format, the
backward compatibity should no longer be needed.

Fix `minimum` and `maximum` on non numeric column

Deprecate passing a column to `type_cast`

I supposed all aggregation functions will return numeric result in
#39039, but that assumption was incorrect for `minimum` and `maximum`,
if an aggregated column is non numeric type.

I've restored type casting aggregated result for `minimum` and `maximum`.

Fixes #39110.

The type information for type casting is entirely separated to type
object, so if anyone does passing a column to `type_cast` in Rails 6,
they are likely doing something wrong. See the comment for more details:

https://github.com/rails/rails/blob/28d815b89487ce4001a3f6f0ab684e6f9c017ed0/activerecord/lib/active_record/connection_adapters/abstract/quoting.rb#L33-L42

This also deprecates passing legacy binds (an array of `[column, value]`
which is 4.2 style format) to query methods on connection. That legacy
format was kept for backward compatibility, instead of that, I've
supported casted binds format (an array of casted values), it is easier
to construct binds than existing two binds format.

Lock benchmark-ips version < 2.8

Looks like "File does not exist: benchmark/ips/stats/stats_metric" is
caused by benchmark-ips 2.8.0.

https://buildkite.com/rails/rails/builds/68586#d77847d6-b3f5-4fd9-af9e-45f465c9d801/940-965
https://buildkite.com/rails/rails/builds/68586#e2037c59-a695-41ff-9040-c8ee74fa7c16/940-965

Looks like benchmark-ips 2.8.0 is broken.

https://buildkite.com/rails/rails/builds/68583#cefd24fb-33e5-4967-ab5d-f52c46700cd7/1879-1934

Remove internal `without_transaction_enrollment` callbacks

Avoid quite useless loop in `find_by_sql`

[Doc] Add mention to shared section in `config_for` docs [ci skip]

We were testing the behavior of out of range ID's on where queries but
not when out of range slugs are converted to ID's.

Also batch attribute readers and writers

I've found the internal `without_transaction_enrollment` callbacks which
have not been newly used over five years, when I tried to work reverting
#9068 (https://github.com/rails/rails/pull/36049#issuecomment-487318060).

I think that we will never make that callbacks public, since the
mechanism of `without_transaction_enrollment` is too implementation
specific, at least before #9068, records in a transaction had enrolled
all into the transaction.

That callbacks was introduced at #18936 to make `touch_later` #19324,
but I think that the internal callbacks is overkill to just make the
`touch_later` only, and invoking the extra callbacks also have a little
overhead even if we haven't used that.

So I think we can remove the internal callbacks for now, until we will
come up with a good use for that callbacks.

37913 added the possibility to deeply merge configurations by grouping
them within a shared section. This powerful alternative was not reflected
in any documentation, which made my team think it was not possible until
I found out this feature after looking at the source code.

This patch reflects this change in the documentation so that it is
easier for other developers to know about this behavior.

`column_types` is empty except PostgreSQL adapter, and
`attribute_types.each_key { |k| column_types.delete k }` is also empty
even if PostgreSQL adapter almost all case, so that code is quite
useless. This improves performance for `find_by_sql` to avoid that
useless loop as much as possible.

```ruby
ActiveRecord::Schema.define do
  create_table :active_storage_blobs do |t|
    t.string   :key,          null: false
    t.string   :filename,     null: false
    t.string   :content_type
    t.text     :metadata
    t.string   :service_name, null: false
    t.bigint   :byte_size,    null: false
    t.string   :checksum,     null: false
    t.datetime :created_at,   null: false

    t.index [ :key ], unique: true
  end
end

class ActiveStorageBlob < ActiveRecord::Base
end

Benchmark.ips do |x|
  x.report("find_by") { ActiveStorageBlob.find_by(id: 1) }
end
```

Before:

```
Warming up --------------------------------------
             find_by     1.256k i/100ms
Calculating -------------------------------------
             find_by     12.595k (± 3.4%) i/s -     64.056k in   5.091599s
```

After:

```
Warming up --------------------------------------
             find_by     1.341k i/100ms
Calculating -------------------------------------
             find_by     13.170k (± 3.5%) i/s -     67.050k in   5.097439s
```

To avoid column types loop for PostgreSQL adapter, this skips returning
additional column types if a column has already been type casted by pg
decoders. Fortunately this fixes #36186 partly for common types.

https://buildkite.com/rails/rails/builds/68559#2decc846-c6b5-46f7-a805-426a8063d36c/1016-1027

Remove useless type cast on aggregated value

Update pg gem required version to 1.1

Relying on the `Arel::Table.engine` is convenient if an app have only a
single kind of database, but if not so, the global state is not always
the same with the current connection.

Batch attribute methods definition

Add override of unary plus for `ActiveSupport::Duration`

Deprecate `allowed_index_name_length` in `DatabaseLimits`

```ruby
(+ 1.second).class
=> ActiveSupport::Duration
```

Fixes #39079.

The less we depend on require_dependency, the better.

`allowed_index_name_length` was used for internal temporary operations
in SQLite3, since index name in SQLite3 must be globally unique and
SQLite3 doesn't have ALTER TABLE feature (so it is emulated by creating
temporary table with prefix).

`allowed_index_name_length` was to reserve the margin for the prefix,
but actually SQLite3 doesn't have a limitation for identifier name
length, so the margin has removed at 36901e6e.

Now `allowed_index_name_length` is no longer relied on by any adapter,
so I'd like to remove the internal specific method which is no longer
used.