Keep in sync with 3-2-stable.

I merged b8837066 but forgot to move the entry past beta1.

1. When comparing the directory to delete against the top level
   cache_path, use File.realpath to make sure we aren't comparing two
   unequal strings that point to the same path. This occurs, for
   example, when cache_path has a trailing slash, which it does in the
   default Rails configuration. Since the input to
   delete_empty_directories never has a trailing slash, the comparison
   will never be true and the top level cache directory (and above) may
   be deleted. However…

2. File.delete raises EPERM when trying to delete a directory, so no
   directories have ever been deleted. Changing the code to Dir.delete
   fixes that.

This reverts commit 046e27a7.
Check https://github.com/rails/rails/pull/8815#issuecomment-14026122 for
details.

Reason: on a second thought, minitest prints a seed as run option
regardless of the test order, and it actually calls srand, so
albeit it might be misleading I believe, it is the way it is and
should be left that way.

This reverts commit c15862ae.

See the first FIXME comment in the patch for the rationale.

Since all changes from #9347 are related to AS, it seems proper that the
test is placed there as well.

ActiveSupport::NumberHelper#number_to_human now returns the number unaltered when
the units hash does not contain the needed key, e.g. when the number provided is less
than the largest key provided.

In i18n gem, the following methods are defined.
- `Hash#except`
- `Hash#slice`

But if there are defined already, i18n skips these definitions.
So these definition by `active_support` are required before `require 'i18n'`.

In `i18n` gem, the method definition of  `Hash#deep_merge!` is skipped
when it is already defined.

- warning: method redefined; discarding old deep_merge!

Release notes at: http://www.unicode.org/versions/Unicode6.2.0/

added test for compress options of gzip

update changelog

And improve AS changelog a bit [ci skip]

Modify TimeWithZone#as_json to return 3DP of sub-second accuracy by default, since it's allowed by the spec and is very useful.

Minitest has changed to require this since https://github.com/seattlerb/minitest/commit/728054b

[ci skip]

this would give us some more clues in case a test silently dies inside Thread