Maybe we can eliminate group committer thread

[zz-jason]

Enhancement

Currently, the group committer thread is used for:

1. Calculate the minimal global user and system transaction IDs, determining whether a user transaction with remote dependency can be committed safely.
2. Collect and flush WAL records from all worker threads to a single WAL file

The proposed solution is:

1. remove group committer
2. each worker writes its WAL records to the local WAL file, updates the flushed system and user transaction id on the worker context
3. for user transactions with remote dependency, collect flushed system and user transaction id from other workers, calculate the global min flushed user and system transaction id, and commit the user transaction if safe
4. for recovery, read all the wall files, shuffle by page id, sort by page sequence number in the wall record, and replay the wal records
5. for checkpoint, the global user and system transaction id can be used to indicate lower watermarks and to truncate wal files

[zz-jason]

A simple POC on ycsb workload A and B shows scalable performance improvements as worker thread increased, and great performance improvement over the group committer.