When trying to access a character on a string buffer object via `:[]`, if the object being accessed currently returns `html_safe?` as true,
we used to set  `@html_safe` variable as true on new object created. When doing something like

x = 'Hello'.html_safe
x[/a/, 1]

would throw an error on ruby 2.2, since when nothign gets matched nil is returned by the code and it tries to set   `@html_safe` value to true,
which would error since starting 2.2 nil is frozen.

This change adds a safety net to avoid setting `@html_safe = true` on frozen objects.

Fixes #18235

It should be part of the documented public API, since we have an entire
section of the guides dedicated to it. Documented in a way that
addresses the concerns which kept it undocumented in the past.

(This is a manual merge of #9102)

before this change, we were allocating AS::SafeBuffer objects that were
being interpolated in to a string, so the safe buffer object was being
thrown away.  This change only allocates a string (vs a string *and* a
safebuffer) and interpolates the string.

On my test application, this reduced the AS::SafeBuffer objects from
1527k per request to about 500 per request.

html_escape_interpolated_argument is only used in mutation methods:

  https://github.com/rails/rails/blob/c07d09559ec171e1904b55c7ad7e8c7d586ca51b/activesupport/lib/active_support/core_ext/string/output_safety.rb#L174
  https://github.com/rails/rails/blob/c07d09559ec171e1904b55c7ad7e8c7d586ca51b/activesupport/lib/active_support/core_ext/string/output_safety.rb#L179

The return value doesn't need to be converted to an AS::SafeBuffer since
we know that the current object is an AS::SafeBuffer and will be
mutated, and the return value from html_escape_interpolated_argument
will be thrown away

Change to require all active_support/deprecation since that's the actual
entry point for the deprecation methods.

Make `#prepend` method modify instance in-place and return self
instead of just returning modified value. That is exactly what
`#prepend!` method was doing previously, so it's deprecated from
now on.

The behavior of json_escape was fixed in 2f1c5789, but the doc
changes and example in that commit incorrectly indicated that the
return value would be html-safe. Since quotation marks are
preserved, the raw value is not safe to use in other contexts
(specifically HTML attributes).

Interpolation was untested and did not work with hash arguments.

Adds
- support for interpolation with hash argument
- tests for the above
- tests for safe/unsafe interpolation

Use the already existing strings instead of creating a new one each time
just to test if it responds to the methods.

This reverts commit 8ce3c1e5, reversing
changes made to f93da579.

Reason: It slow down the running time.

require "diffbench"
load 'output_safety.rb'

N = 10000
b = ActiveSupport::SafeBuffer.new("hello world")
DiffBench.bm do
  report "capitalize in safe buffer" do
    N.times do
      b.capitalize
    end
  end
end

> git checkout  069ea45c; diffbench bench.rb;
diffbench bench.rb;diffbench
bench.rb;diffbench bench.rb;diffbench
bench.rb;diffbench bench.rb;diffbench
bench.rb;

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.009733)
Before patch:   0.010000   0.000000   0.010000
(  0.007702)
Improvement: -26%

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.009768)
Before patch:   0.010000   0.000000   0.010000
(  0.007896)
Improvement: -24%

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.009938)
Before patch:   0.010000   0.000000   0.010000
(  0.007768)
Improvement: -28%

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.010001)
Before patch:   0.010000   0.000000   0.010000
(  0.007873)
Improvement: -27%

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.009670)
Before patch:   0.010000   0.000000   0.010000
(  0.007800)
Improvement: -24%

Running benchmark with current working tree
Checkout HEAD^
Running benchmark with HEAD^
Checkout to previous HEAD again

                    user     system      total
                    real
----------------------------------capitalize
in safe buffer
After patch:    0.010000   0.000000   0.010000
(  0.009949)
Before patch:   0.010000   0.000000   0.010000
(  0.007752)
Improvement: -28%

https://www.owasp.org/index.php/XSS_%28Cross_Site_Scripting%29_Prevention_Cheat_Sheet#RULE_.231_-_HTML_Escape_Before_Inserting_Untrusted_Data_into_HTML_Element_Content
Closes #7215

This reverts commit fa6d921e.

Reason: Not a fan of such massive changes. We usually close such changes
if made to Rails master as a pull request. Following the same principle
here and reverting.

[ci skip]

For future reference, this is the regex I used: ^\s*#\s*\n(?!\s*#). Replace
with the first match, and voilà! Note that the regex matches a little bit too
much, so you probably want to `git add -i .` and go through every single diff
to check if it actually should be changed.

Logic in clone_empty method was dealing with old @dirty variable, which
has changed by @html_safe in this commit:
https://github.com/rails/rails/commit/139963c99a955520db6373343662e55f4d16dcd1

This was issuing a "not initialized variable" warning - related to:
https://github.com/rails/rails/pull/5237

The logic applied by this method is already handled by the [] override,
so there is no need to reset the variable here.