The Problem
-----------

The following line from `String#camelize`:

  string = string.sub(/^(?:#{inflections.acronym_regex}(?=\b|[A-Z_])|\w)/) { |match| match.downcase }

and the following line from `String#camelize`:

  word.gsub!(/(?:(?<=([A-Za-z\d]))|\b)(#{inflections.acronym_regex})(?=\b|[^a-z])/) { "#{$1 && '_'.freeze }#{$2.downcase}" }#{$2.downcase}" }

Both generate the same regexep in the first part of the `.sub`/`.gsub`
method calls every time the function is called, creating an extra object
allocation each time.  The value of `acronym_regex` only changes if the
user decides add an acronym to the current set of inflections and apends
another string on the the regexp generated here, but beyond that it
remains relatively static.

This has been around since acronym support was introduced back in 2011
in PR#1648.

Proposed Solution
-----------------
To avoid re-generating these strings every time these methods are
called, cache the values of these regular expressions in the
`ActiveSupport::Inflector::Inflections` instance, making it so these
regular expressions are only generated once, or when the acronym's are
added to.

Other notable changes is the attr_readers are nodoc'd, as they shouldn't
really be public APIs for users.  Also, the new method,
define_acronym_regex_patterns, is the only method in charge of
manipulating @acronym_regex, and initialize_dup also makes use of that
new change.

** Note about fix for non-deterministic actionpack test **

With the introduction of `@acronym_underscore_regex` and
`@acronym_camelize_regex`, tests that manipulated these for a short
time, then reset them could caused test failures to happen.  This
happened because the previous way we reset the `@acronyms` and
`@acronym_regex` was the set them using #instance_variable_set, which
wouldn't run the #define_acronym_regex_patterns method.

This has now been introduced into the actionpack tests to avoid this
failure.