Fix OOM caused by queue pressure

[AdamKatzDev]

Fixes #307. Should be replaced in favor of #262 (but my PR might be used to fix the Kafka sink).

I was able to get stable performance during snapshots and huge single statement updates:

1. Used LinkedBlockingQueue (which is also thread-safe) with limit of 200,000 records instead of ConcurrentLinkedQueue. When a queue becomes full thread sleeps in a loop until space in the queue is freed. I didn't use @IlyaTsoi solution since ConcurrentLinkedQueue::size is an O(n) (i.e. quite slow for our purpose).
2. Moved records to a separate buffer with fixed size of 100k to prevent code from building a single batch indefinitely.

What I don't like here:

1. Busy-waiting in attempt to put a record in an already full queue.
2. Code that moves records to a separate buffer is quite ugly, the implementation might not be optimal for some cases.
3. Lots of magic constants and conditions to make it stable.

I've just recently stumbled upon #262 which in theory should be a perfect solution for the problem and also fixes #342. Alas the PR only solves the problems for the lightweight sink. That is not a problem for me, but looks like the Kafka sink suffers from the same problems, the problematic parts of the code are the same.