Lucene partition balancing: optimize and fix transaction timeouts

[jjezra]

Currently, the rebalancing code uses a single transaction to iterate through the grouping keys and rebalance the first partition in need in each group. This may cause a transaction timeout, and limits the number of documents that can moved - which causes extra merges.
The suggested solution is to:

1. Use a read-only Agility Context to read the grouping keys and partitions info
2. Use the same AC to evaluate if a rebalancing is needed on a give partition
3. If needed, create a new, dedicated transaction to rebalance each single partition