This PR adds the allocations to the instrumentation for template and
partial rendering.

Before:

```
  Rendering posts/new.html.erb within layouts/application
  Rendered posts/_form.html.erb (9.7ms)
  Rendered posts/new.html.erb within layouts/application (10.9ms)
Completed 200 OK in 902ms (Views: 890.8ms | ActiveRecord: 0.8ms)
```

After:

```
  Rendering posts/new.html.erb within layouts/application
  Rendered posts/_form.html.erb (Duration: 7.1ms | Allocations: 6004)
  Rendered posts/new.html.erb within layouts/application (Duration: 8.3ms | Allocations: 6654)
Completed 200 OK in 858ms (Views: 848.4ms | ActiveRecord: 0.4ms | Allocations: 1539564)
```

Since Rails 6.0 will support Ruby 2.4.1 or higher
`# frozen_string_literal: true` magic comment is enough to make string object frozen.
This magic comment is enabled by `Style/FrozenStringLiteralComment` cop.

* Exclude these files not to auto correct false positive `Regexp#freeze`
 - 'actionpack/lib/action_dispatch/journey/router/utils.rb'
 - 'activerecord/lib/active_record/connection_adapters/sqlite3_adapter.rb'

It has been fixed by https://github.com/rubocop-hq/rubocop/pull/6333
Once the newer version of RuboCop released and available at Code Climate these exclude entries should be removed.

* Replace `String#freeze` with `String#-@` manually if explicit frozen string objects are required

 - 'actionpack/test/controller/test_case_test.rb'
 - 'activemodel/test/cases/type/string_test.rb'
 - 'activesupport/lib/active_support/core_ext/string/strip.rb'
 - 'activesupport/test/core_ext/string_ext_test.rb'
 - 'railties/test/generators/actions_test.rb'

In Ruby 2.3 or later, `String#+@` is available and `+@` is faster than `dup`.

```ruby
# frozen_string_literal: true

require "bundler/inline"

gemfile(true) do
  source "https://rubygems.org"

  gem "benchmark-ips"
end

Benchmark.ips do |x|
  x.report('+@') { +"" }
  x.report('dup') { "".dup }
  x.compare!
end
```

```
$ ruby -v benchmark.rb
ruby 2.5.1p57 (2018-03-29 revision 63029) [x86_64-linux]
Warming up --------------------------------------
                  +@   282.289k i/100ms
                 dup   187.638k i/100ms
Calculating -------------------------------------
                  +@      6.775M (± 3.6%) i/s -     33.875M in   5.006253s
                 dup      3.320M (± 2.2%) i/s -     16.700M in   5.032125s

Comparison:
                  +@:  6775299.3 i/s
                 dup:  3320400.7 i/s - 2.04x  slower

```

Two implemented but undocumented features are to help indicate that cache is fresh for 3 hours, and it may continue to be served stale for up to an additional 60 seconds to parallel requests for the same resource or up to 5 minutes while errors are being returned back while the initial synchronous revalidation is attempted.

When running with code triage and derailed benchmarks and focusing on this file:

Before

     16199  /Users/rschneeman/Documents/projects/rails/actionpack/lib/action_controller/metal/strong_parameters.r

After

      2280  /Users/rschneeman/Documents/projects/rails/actionpack/lib/action_controller/metal/strong_parameters.rb

* Call block to #redirect_to in controller context

The documentation for ActionController::Redirecting states that a Proc
argument "will be executed in the controller's context."  However,
unless #instance_eval is used (removed in 6b3ad0ca), that statement is
false for procs defined outside of the controller instance.

This commit restores the documented behavior.

Fixes #33731.

* Move test proc into a constant in another class

Per @rafaelfranca's suggestion.

[Steven Peckins + Rafael Mendonça França]

This updates the `respond_to` method to be code formatted rather than
plain text (as it refers to the method)

The regexp was introduced in 186ac4cd,
and looks cosmetic. While they should be functionally identical in
theory, in practice, case insensitive (but preserving) filesystems can
give results that are differently-cased from the pattern we supplied.

I don't know how to force the filesystem to do the surprising thing,
even when running in an environment that _could_, so no new test.

- Layout/TrailingWhitespace

```
actionpack/lib/action_controller/metal/request_forgery_protection.rb:49:4:
C: Layout/TrailingWhitespace: Trailing whitespace detected.
  #
   ^
```

Related to c3787494

- Performance/StartWith

```
tasks/release.rb:108:44: C: Performance/StartWith:
Use String#start_with? instead of a regex match anchored to the beginning of the string.
      header += "*   No changes.\n\n\n" if current_contents =~ /\A##/
```

I was attempting to use the `flash` functionality in a `Metal`
controller. When including the `flash` concern I received the following
error:

    NoMethodError: undefined method `helper_method'....

Either:

- `AbstractController::Helpers` should be a dependency of
  `ActionController::Flash`
- `ActionController::Flash` should not require the existence of
  `AbstractController::Helpers`.

Since my use case (set a flash and redirect) has no need for the helper
method and that is a common use case, making the dependency conditional
seemed the better option.

NOTE: This is similar to issue #21067 only the error is within Rails
itself while that issue had the error within Devise.

Nesting respond_to calls can lead to unexpected behavior, so it should be
avoided. Currently, the first respond_to format match sets the content-type
for the resulting response. But, if a nested respond_to occurs, it is possible
to match on a different format. For example:

    respond_to do |outer_type|
      outer_type.js do
        respond_to do |inner_type|
          inner_type.html { render body: "HTML" }
        end
      end
    end

Browsers will often include */* in their Accept headers. In the above example,
such a request would result in the outer_type.js match setting the content-
type of the response to text/javascript, while the inner_type.html match will
cause the actual response to return "HTML".

This change tries to minimize potential breakage by only raising an exception
if the nested respond_to calls are in conflict with each other. So, something
like the following example would not raise an exception:

    respond_to do |outer_type|
      outer_type.js do
        respond_to do |inner_type|
          inner_type.js { render body: "JS" }
        end
      end
    end

While the above is nested, it doesn't affect the content-type of the response.

http links will be redirected to the https version, but still better to
just directly link to the https version.

Use attr_reader/attr_writer instead of methods

method is 12% slower

Use flat_map over map.flatten(1)

flatten is 66% slower

Use hash[]= instead of hash.merge! with single arguments

merge! is 166% slower

See https://github.com/rails/rails/pull/32337 for more conversation

Complete renderer documentation

Fixes #28484

Follow up #33401, 5491f811.

Follow up to 9f152a60

- There was an issue inside controller tests where order params were not respected, the reason
  was because we were calling `Hash#to_query` which sorts the results lexicographically.
  1e4e1b62 fixed that issue by not using `to_query` but instead a utility function provided by rack.
- However with the fix came another issue where it's now no longer possible to do this

  ```
   post :foo, params: { user: User.first }

   # Prior to the patch the controller will receive { "user" => "1" }
   # Whereas now you get { "user": "#<User: ...>" }
  ```

  The fix in this PR is to modify `Hash#to_query` to sort only when it
  doesn't contain an array structure that looks something like "bar[]"

  Ref https://github.com/rails/rails/pull/33341#issuecomment-404039396

* Convert hashes into parameters

Ensure `ActionController::Parameters#transform_values` and
`ActionController::Parameters#transform_values!` converts hashes into
parameters.

* fixup! Convert hashes into parameters

[Rafael Mendonça França + Kevin Sjöberg]

From <https://tools.ietf.org/html/rfc5861>:

> The stale-if-error HTTP Cache-Control extension allows a cache to
> return a stale response when an error -- e.g., a 500 Internal Server
> Error, a network segment, or DNS failure -- is encountered, rather
> than returning a "hard" error.  This improves availability.
>
> The stale-while-revalidate HTTP Cache-Control extension allows a
> cache to immediately return a stale response while it revalidates it
> in the background, thereby hiding latency (both in the network and on
> the server) from clients.

These are useful, fully standardized parts of the HTTP protocol with
widespread support among CDN vendors. Supporting them will make it
easier to utilize reverse proxies and CDNs from Rails.

`to_query` sorts parameters before encoding them.  This causes a round
tripping issue as noted here:

  https://github.com/rails/rails/issues/23997#issuecomment-328297933
  https://github.com/rails/rails/issues/10529#issuecomment-328298109
  https://github.com/rails/rails/pull/30558

Unfortunately, that method is being used to generate cache keys, so its
results need to be stable:

  https://github.com/rails/rails/commit/10dec0e65e1f4d87f411b4361045eba86b121be9

However, the test harness is only using `to_query` to encode parameters
before sending them to the controller so the "cache key" usecase doesn't
apply here.

This commit adds a test that demonstrates the round trip problems and
changes the serialization strategy to use Rack for encoding the
parameters rather than `to_query`.

Benchmark:

```ruby
require 'benchmark'
require 'benchmark/ips'
require 'securerandom'

def xor_byte_strings(s1, s2) # :doc:
  s2_bytes = s2.bytes
  s1.each_byte.with_index { |c1, i| s2_bytes[i] ^= c1 }
  s2_bytes.pack("C*")
end

def xor_byte_strings_new(s1, s2) # :doc:
  s2 = s2.dup
  size = s1.bytesize
  i = 0
  while i < size
    s2.setbyte(i, s1.getbyte(i) ^ s2.getbyte(i))
    i += 1
  end
  s2
end

s1 = SecureRandom.random_bytes(32)
s2 = SecureRandom.random_bytes(32)

Benchmark.ips do |x|
  x.report("current"){xor_byte_strings(s1, s2)}
  x.report("new"){xor_byte_strings_new(s1, s2)}
  x.compare!
end

100000.times do |i|
  s3 = SecureRandom.random_bytes(32)
  s4 = SecureRandom.random_bytes(32)
  raise unless xor_byte_strings(s3, s4) == xor_byte_strings_new(s3, s4)
end
```

Results on ruby 2.5.1:

```
Warming up --------------------------------------
             current     6.519k i/100ms
                 new    10.508k i/100ms
Calculating -------------------------------------
             current     84.723k (_ 0.4%) i/s -    423.735k in   5.001456s
                 new    145.871k (_ 0.3%) i/s -    735.560k in   5.042606s

Comparison:
                 new:   145870.6 i/s
             current:    84723.4 i/s - 1.72x  slower
```

Make it clear that the return value is converted to an
instance of ActionController::Parameters if possible

The example code is meant to be a string.

If a POST request is followed by a GET request in a controller test, the
`rack.input` and `RAW_POST_DATA` headers from the first request will be
reset but the `CONTENT_LENGTH` header will leak, leading the request
object in the second request to incorrectly believe it has a body.

`RAW_POST_DATA` is derived from the `rack.input` header, which changes
with each test request. It needs to be cleared in `scrub_env!`, or all
requests within the same test will see the value from the first request.

`permit!` is intended to mark all instances of `ActionController::Parameters` as permitted, however nested arrays of params were not being marked permitted because the method did shallow iteration.
This fixes that by flattening the array before calling `permit!` on all each item.