Match core flushing logic and ensure expiration always happens

In the current implementation, garbageCollection() performs expiration based on variable cache_flush_$bin being non zero, expecting it to represent a timestamp. However, immediately after noticing a flush is in order, it resets that variable to zero to avoid a flush stampede.

This variable now being 0, the next garbage collections will no longer flush anything until that variable is reset to non-zero, which only happens when clear(NULL, $bin) is invoked (see MongoDBCache::clear()), typically via cache_clear_all(NULL, $bin), or cache_clear_all() on its own for core bins.

However, drupal_flush_all_caches() invokes cache_clear_all('*', $table, TRUE), meaning even a full cache clear will not trigger a garbage collection, but just truncate the core predefined collections and those defined by hook_flush_caches() implementations, and not re-enable the expiration cycle.

Workaround: manually invoke these clears, like drush ev 'cache_clear_all(NULL, "some_bin");' for each bin.

Suggested fix:

• also reset cache_flush_$bin, in the $cid = '*', $wildcard = TRUE case.
• implement mongodb_cache_flush_caches() to declare the bins, but this raises the question of identifying them: just relying on the collection name may not be enough, as some developers may name cache collections with names not matching cache_*, or may name non-cache collections with names matching this pattern. A solution could be to add a collection prefix so that the collection name could be relied upon to find cache collections.
• implement #1802114: Take advantage of TTL collections which would let the server perform the cleanup itself, and would save us having to query the DB in garbageCollection()