Hashes and leaves can be written to disk on multiple threads

`MerkleDbDataSource.saveRecords()` splits the process into three parts:

* Write hashes to "hash chunks store"
* Write leaves to "leaves store"
* Write key to path mappings to "keys store"

The third one is heavily multi-threaded. It includes loading object key buckets, updating them, and then writing to disk, all implemented as separate tasks with dependencies. However, the first and the second parts are both single-threaded. They become the bottleneck. All keys/path mappings have been already updated, there are threads to utilize, but hashes and leaves are still being written. It increases total flush time significantly.