Increments in an existing Transaction may populate Memcache incorrectly

[sappenin]

Steps to recreate:

• Start a Datastore TX.
• Increment a counter that is not currently in memcache, setting "isolated" to false so that the increment occurs as part of the current TX.
• ShardedCounterServiceImpl#increment will end with a call to #getCounter, which will query the datastore for Shard values in an attempt to get the counter's count. Since the initial TX is still open, no shards will be returned (any loads from the Datastore will reflect the current state of the datastore as of the beginning of the TX).
• Memcache will be initialized to 0, which is should really be 1 or some other value that's truly the count. Or, if the counter has "count", but hasn't been accessed in a while, then an old value of the count may be populated into memcache.

This problem will go away after the cache entry is evicted, but can cause an incorrect counter count for a short period of time.

Possible Fixes:

1. Remove the return value of all increment methods so the cache is not updated until all TX's have concluded.