Fix/111

[numberoverzero]

This fixes a bug where the records of the last GetRecords call in a closed Shard are dropped from the Coordinator's buffer immediately. See #111 for investigation details.

Now, the Coordinator keeps a count of buffered records for each closed shard that it previously tracked; as each record is popped off the buffer, the count is decremented until it's dropped. The Shards can't be tracked in the same place as active shards (Coordinator.active) because those are periodically polled for more records. However, we can't stop tracking them because then the rendered Shard.token would skip ahead while the records are still buffered.

In addition to unit tests, I ran some hacked together tests for a few hours; records were inserted every 30mins/1min/10sec in 3 different streams and then counted on the way out of the Stream processor.