Fixes #23524

Having collection caching that wraps templates and automatically tries
to infer if they are cachable proved to be too much of a hassle.

We'd rather have it be something you explicitly turn on.

This removes much of the code and docs to explain the previous automatic
behavior.

This change also removes scoped cache keys and passing cache_options.

Creating a protected getter method for `@parameters`.

While iterating an AC::Parameters object, the object will mutate itself
and stick AC::Parameters objects where there used to be hashes:

  https://github.com/rails/rails/blob/f57092ad728fa1de06c4f5fd9d09dcc2c4738fd9/actionpack/lib/action_controller/metal/strong_parameters.rb#L632

If you use `permit` after this iteration, the `fields_for_style` method
wouldn't return true because the child objects are now AC::Parameters
objects rather than Hashes.

fixes #23701

Now that AC::Parameters is no longer a Hash, it shouldn't look like a hash.

The `cookies` hash isn't updated with the value generated by the output
from `to_header` so it wasn't testing anything. Rendering the cookie
value in the controller makes sure that the escaping is actually working.

`NEVER_UNPERMITTED_PARAMS` is deprecated in Rails 4.2. See #15933.

A callable cache key writes to the collection cache under a certain namespace.
Which means if we don't have scoped cache key we can just rely on the
`cache model_name do` in the templates to cache them.

Less writes, more sharing.

Add `assert_customer_cached` to better illustrate this in tests, and remove
tests which then don't communicate as much.

It was difficult to see when the partials were rendered, and how many times
we expected it to be rendered before. Because we weren't explaining it.

When people pass `cache: -> item { item.upcase }` they scope the collection
cache keys so the individual partial cache isn't reused.

Test that behavior.

We're not guaranteed to have a `RequestEncoder` to assign on `get` requests
because we aren't extracting the parser from the response content type.

Until now.

When testing:

```ruby
post articles_path, params: { article: { title: 'Ahoy!' } }, as: :json
```

It's common to want to make assertions on the response body. Perhaps the
server responded with JSON, so you write `JSON.parse(response.body)`.
But that gets tedious real quick.

Instead add `parsed_body` which will automatically parse the reponse
body as what the last request was encoded `as`.

Tests can (and do) access the database from the main thread.  In this
case they were starting a transaction, then making a request.  The
request would create a new thread, which would allocate a new database
connection.  Since the main thread started a transaction that contains
data that the new thread wants to see, the new thread would not see it
due to data visibility from transactions.  Spawning the new thread in
production is fine because middleware should not be doing database
manipulation similar to the test harness.  Before 603fe20c it was
possible to set the database connection id based on a thread local, but
603fe20c changes the connection lookup code to never look at the
"connection id" but only at the thread object itself.  Without that
indirection, we can't force threads to use the same connection pool as
another thread.

Fixes #23483

There are some cases when the `body` in `response_body=` can be set to
nil. One of those cases is in `actionpack-action_caching` which I found
while upgrading it for Rails 5.

It's not possible to run `body.each` on a `nil` body so we have to
return after we run `response.reset_body!`.

This removes the following warning.

```
rails/actionpack/test/controller/render_test.rb:278: warning: assigned but unused variable - response
```

This works on OSX but for some reason travis is throwing a
```
  1) Error:
ExpiresInRenderTest#test_dynamic_render_with_absolute_path:
NoMethodError: undefined method `unlink' for nil:NilClass
```
Looking at other tests in Railties the file has a name and we close
it before unlinking, so I'm going to try that.

Test that we are not allowing you to grab a file with an absolute path
outside of your application directory. This is dangerous because it
could be used to retrieve files from the server like `/etc/passwd`.

[ci skip]

The cache for `render file:` seems to also be used in the case of
`render(string)`.  If one is supposed to be a hit and the other is
supposed to be a miss, and they both reference the same file, then the
cache could return incorrect values.  This commit clears the cache
between runs so that we get non-cached behavior.

CVE-2016-0752

See bug #21032.

- Fix warning introduced in https://github.com/vipulnsward/rails/commit/77acc004efad07dfd4d4f83be14ef897968a3fd9 when fixing API  responses.
- render :text is deprecated, so added an assertion for it.

- Fixes bug #23142.
- Bug was occurring only with ActionController::API, because `_process_options` wasn't being run for API requests, even though it was being run for normal app requests.

This reverts commit 5d1b7c3b.

The change here didn't actually fix the issue it was trying to fix, and
this isn't the correct way to fix either issue. The problem is switching
from the builder to grouping with find_all/regex is now very dependent
on how you structure your path pattern.

Response etags to always be weak: Prefixed W/ to value returned by ActionDispatch::Http::Cache::Response#etag= such that etags set in fresh_when and stale? are weak. For #17556.

It's reasonable to expose different value readers.

We can provide a more flexible upgrade experience by warning users they are using unsafe methods instead of forcing the safe API by deprecating before removal. This PR provides this functionality.

In commit d993cb36 `build_path` was changed from using `grep` to
`find_all` to save array allocations.

This change was a little too aggressive in that when the dash comes
before the symbol like `/omg-:song` the symbol is skipped.

Removing the check for `n.right.left.literal?` fixes this issue, but
does add back some allocations. The number of allocations are still well
less than before.

I've added a regression test to test this behavior for the future.

Fixes #23069.

Array allocations as of d993cb36:

```
{:T_SYMBOL=>11}
{:T_REGEXP=>17}
{:T_STRUCT=>6500}
{:T_MATCH=>12004}
{:T_OBJECT=>91009}
{:T_DATA=>100088}
{:T_HASH=>114013}
{:T_STRING=>159637}
{:T_ARRAY=>321056}
{:T_IMEMO=>351133}
```

Array allocations after this change:

```
{:T_SYMBOL=>11}
{:T_REGEXP=>1017}
{:T_STRUCT=>6500}
{:T_MATCH=>12004}
{:T_DATA=>84092}
{:T_OBJECT=>87009}
{:T_HASH=>110015}
{:T_STRING=>166152}
{:T_ARRAY=>322056}
{:T_NODE=>343558}
```

Fixes #23026

See discussion at #23026

Turns

```
post articles_path(format: :json), params: { article: { name: 'Ahoy!' } }.to_json,
  headers: { 'Content-Type' => 'application/json' }
```

into

```
post articles_path, params: { article: { name: 'Ahoy!' } }, as: :json
```