Currently, if `:datetime` type has a precision, type casting always does
round off subseconds regardless of whether necessary or not, it is a bit
slow.

Since #34970, `t.timestamps` with `precision: 6` by default, so
`pluck(:created_at)` in newly created app will always be affected by the
round off.

This avoids the round off if possible, it makes `pluck(:created_at)`
about 25% faster.

https://gist.github.com/kamipo/e029539f2632aee9f5e711fe66fc8842

Before (0a87d7c9 with postgresql adapter):

```
Warming up --------------------------------------
    users.pluck(:id)      344.000  i/100ms
  users.pluck(:name)      336.000  i/100ms
users.pluck(:created_at)  206.000  i/100ms
Calculating -------------------------------------
    users.pluck(:id)      3.620k (± 8.5%) i/s -     18.232k in   5.077316s
  users.pluck(:name)      3.579k (± 9.4%) i/s -     17.808k in   5.020230s
users.pluck(:created_at)  2.069k (± 8.0%) i/s -     10.300k in   5.019284s
```

Before (0a87d7c9 with mysql2 adapter):

```
Warming up --------------------------------------
    users.pluck(:id)      245.000  i/100ms
  users.pluck(:name)      240.000  i/100ms
users.pluck(:created_at)  180.000  i/100ms
Calculating -------------------------------------
    users.pluck(:id)      2.548k (± 9.4%) i/s -     12.740k in   5.066574s
  users.pluck(:name)      2.513k (± 8.0%) i/s -     12.480k in   5.011260s
users.pluck(:created_at)  1.771k (±11.2%) i/s -      8.820k in   5.084473s
```

After (this change with postgresql adapter):

```
Warming up --------------------------------------
    users.pluck(:id)      348.000  i/100ms
  users.pluck(:name)      357.000  i/100ms
users.pluck(:created_at)  254.000  i/100ms
Calculating -------------------------------------
    users.pluck(:id)      3.628k (± 8.2%) i/s -     18.096k in   5.024748s
  users.pluck(:name)      3.624k (±12.4%) i/s -     17.850k in   5.020959s
users.pluck(:created_at)  2.567k (± 7.0%) i/s -     12.954k in   5.081153s
```

After (this change with mysql2 adapter):

```
Warming up --------------------------------------
    users.pluck(:id)      268.000  i/100ms
  users.pluck(:name)      265.000  i/100ms
users.pluck(:created_at)  207.000  i/100ms
Calculating -------------------------------------
    users.pluck(:id)      2.586k (±10.9%) i/s -     12.864k in   5.050546s
  users.pluck(:name)      2.543k (±10.2%) i/s -     12.720k in   5.067726s
users.pluck(:created_at)  2.263k (±14.5%) i/s -     10.971k in   5.004039s
```

Allow symbol (i.e. quoted identifier) as safe SQL string

`pluck(:id)` / `order(:id)` are very common use case, and passed symbol
(i.e. quoted identifier) is obviously safe argument, but
`:id.to_s.split(/\s*,\s*/).all? { |part| permit.match?(part) }` is
useless and a bit expensive operation for each such safe symbols (will
make extra 2 mutable strings, 1 array, 1 proc).

This avoids the expensive operation to such safe symbols, it makes
`pluck(:id)` / `order(:id)` itself about 9% faster.

https://gist.github.com/kamipo/11d428b57f3629a72ae89c6f21721326

Before (93e64073):

```
Warming up --------------------------------------
    users.pluck(:id)     1.217k i/100ms
users.order(:id).to_sql  1.848k i/100ms
Calculating -------------------------------------
    users.pluck(:id)     12.239k (± 8.2%) i/s -     60.850k in   5.013839s
users.order(:id).to_sql  19.111k (± 7.5%) i/s -     96.096k in   5.064450s
```

After (this change):

```
Warming up --------------------------------------
    users.pluck(:id)     1.293k i/100ms
users.order(:id).to_sql  2.036k i/100ms
Calculating -------------------------------------
    users.pluck(:id)     13.257k (± 6.9%) i/s -     65.943k in   5.002568s
users.order(:id).to_sql  20.957k (± 7.6%) i/s -    105.872k in   5.086102s
```

[documentation][ci skip] stringify_keys and symbolize_keys have stable results.

Remove unused requires from debug_exceptions

Rails 6 uses the `Hash.transform_keys` found in Ruby 2.5 and later, and that method enumerates keys based on insertion order.  Calling `symbolize_keys`, `stringify_keys`, and their bang variants will result in the same hash every time -- the value for any key where a collision occurs is the last assigned in that enumeration

In the docs for Hash --  https://ruby-doc.org/core-2.5.0/Hash.html

> Hashes enumerate their values in the order that the corresponding keys were inserted.

[documentation][ci skip] changes 'Week day' to 'day of week'

'Week day' has a specific meaning in English -- see https://en.wiktionary.org/wiki/weekday for details -- that is not meant here.  'Day of week' is more appropriate.

[ci skip]

Use WeakRef in Reaper to avoid leaking connection pools

Prior to 3e2e8eeb the Reaper thread
would hold a reference to connection pools indefinitely, preventing the
connection pool from being garbage collected, and also leaking the
Thread.

Since 3e2e8eeb, there is only one Reaper
Thread for all pools, however all pools are still stored in a class
variable, preventing them from being garbage collected.

This commit instead holds reference to the pools through a WeakRef. This
way, connection pools referenced elsewhere will be reaped, any others
will be able to be garbage collected.

I don't love resorting to WeakRef to solve this, but I believe it's the
simplest way to accomplish the the desired behaviour.

Change comments

[ci skip]

Fix EventedFileUpdateChecker through a symlink 

Wrap ActionView cache expiry in a mutex

Delete evented_file_update_checker existing_parent

On MacOS, Dir.tmpdir gives me a folder inside "/var/folders/". However,
/var is a symlink to /private/var.

Previously, the nonexistent directory test would fail because it was
initialized with /var/folders/... but the filenames from listen would be
the realpaths.

This commit normalizes the dirs by calling realpath on them if they
exist. This is done on boot!, so it will work with newly directories
through the symlink.

The common include of this test creates a tmpdir, we should use that for
consistency.

This is no longer used as of caa3cc88

 Bump Gemfile entry jbuilder version

Add :allow_nil option to delegate_missing_to

attachment

jbuilder 2.6.4 is the first version that relaxes the version constraint to allow
Rails 6. I also did some more tests in #25183, although not with 2.6.4
explicitly.

To simplify the version requirement, I went for 2.7.

https://github.com/rails/jbuilder/blob/v2.6.4/jbuilder.gemspec
https://github.com/rails/rails/issues/25183#issuecomment-494342406

Keep part when scope option has value

After @kamipo CR feedback we realized `Route#build` wasn't used. As it
is also private API being able to create Routes both with `#new` and
`#build` was redundant.

This commit changes from constructor's argument list to keyword
arguments in order to remove the dependency of remember parameters'
positions.

It also unifies all parameters extracted from the `scope` into
`scope_params`, which also takes care of providing the default values
for them.

This commit changes from constructor's argument list to keyword
arguments in order to remove the dependency of remember parameters'
positions.

The constructor already provided a default value for `internal`, this
commits takes the chance to also add default values for `precedence` and
`scope_options`.

When a route was defined within an optional scope, if that route didn't
take parameters the scope was lost when using path helpers. This patch
ensures scope is kept both when the route takes parameters or when it
doesn't.

Fixes #33219

Remove compatibility module from docs [ci skip]

This module exists to warn old users. I think we should remove
it from the docs so we don't advertise it.

Added deep_transform_keys to StrongParameters

- This hook was added in 90296674

[CI skip] Put :nodoc: on method that raises NoMethodError