Adds new set_versioned and get_versioned APIs to cache, which means you can
request a specific version from cache and it will not return an outdated
version.

This is important when using multi-layer-caches where a local cache might have
an outdated version of the cache, but the shared cache has a current version.
With this feature, the content of the cache does not have to be rebuilt, as
it will automatically retrieve it from the shared cache if necessary.

For cache types which mean this information can be obtained without a
significant performance cost (i.e. just by calling strlen and not
having to serialize something that wasn't serialized already),
this change calculates the size of data read from or written to cache
in each request and includes it in the perfdebug table at bottom of
output (when that is turned on).

This supports the following cache types:

* File store
* Redis (only if caching is enabled)

Under parallel development, when something is worth being commented
in upgrade.txt notes, it will be applied to all the development branches
using the "earliest" one to be released. So, if something changes
in 3.10 (and also is applied to master), the upgrade.txt notes will
show 3.10 in both branches.

Of course, if something only goes to one dev branch (say master), then the
master version will be used in the notes (4.0 in this case).

The scope of this change is limited to modifying the MongoDB cache
plugin to use the new version of the driver that accepts PHP 7.x
versions. Following the instructions in the official MongoDB
documentation, https://docs.mongodb.com/php-library/current/, the
MongoDB PHP Library was included and used.

Russell Smith authored Jun 07, 2016 and Mark Nelson committed Mar 02, 2017

It is now safe to cache a reference to a cache and expect consistent results.

Changing identifiers altered cache results where a reference was
held to the cache. Identifiers have been set to be cached with
identifiers included so the caches are separate.

As a consequence of this it was identified that invalidation events
and identifiers don't easily work together as an event can't determine
which identifiers should be used for cache invalidation.  So invalidation
events have been made incompatible with identifiers being set.  No core
code used this combination as it's not possible to understand any expected
behaviour.

Event invalidation for application and session caches was centralised to the same
location.  The only difference was the name of the lastinvalidation variable. This
improves support and consistency of invalidation code.

purge_all() and purge_by_definition() look in the configuration
for which caches are available and then creates them to purge them.
The configuration stores the values used by initialise(), not
initialise_unit_test_instance() and would therefore fail to purge
all caches if they were not purged by another means.

In the case of filestore, it's purged by unit tests, in the case
of memcache(d), it purges the whole store when a single definition
is requested.

Therefore all configuration was moved into the configuration file
during unit tests and does not have any special override codes in
the unit test infrastructure.

When added a memcached instance you can now select whether the cache is
being shared by other applications. The setting will determine the
purging strategy.

Shared caches will have individual keys deleted while dedicated caches
will have the entire cache purged (better performance over networks).

Note: This option only works with the correct version of the php
memcached extension and with the multi-site safe changes.

If the cache does all the dereferencing when it stores and loads
the objects, then the cache loaders don't need to do that work.
This is true of all caches that use something other than PHP's memory
to store their results.

The cache store mode is now included in the performance stats
printed at the bottom of the page.
It is represented as either [a] [s] or [r] and a title is used
to actually state the mode.

Cache stores are now responsible for deciding if they are suitable
for use as the primary cache store during unit and acceptance tests

Changes are as follows:
* New method: cache_store::ready_to_be_used_for_testing
* Core cache store have been updated to override this method as
  required.
* MongoDB usesafe default changed to true to match actual default
  behaviour.

We now use the unit test cache config class when running
acceptance tests so that the same defines are applicable there.

Cache data source aggregate functionality was found to be broken
and unused, because of this the decision was made to remove it
rather than fix it.
As it was broken we did not follow typical deprecation methods and
instead the code was removed outright with only structure
remaining and left deprecated.

Sam Hemelryk authored Jan 02, 2014 and Dan Poltawski committed Jan 07, 2014

(Fixes the fixed forwarded static call bug reverted in a5581c73)

This reverts commit 2201c02d.