counters++ coalesce counter mutations with the same partition key

CASSANDRA-6504 counter caching helps with hot counter cells, but doesn't really help when we have a hot counter partition and different cells within being updated simultaneously by different clients (the striped locks are partition-level, not cell-level).

To improve performance in this scenario, we could coalesce the mutations in COUNTER_MUTATION stage if they share the same partition key + target cf, and apply them together, sharing a single lock.

If beneficial, this can go into 2.1.x or 3.0, doesn't have to be 2.1.0.