Currently, we have 2 separate code paths for memtx replace w/ and w/o transaction management, and the transaction manager is taking the responsibility for replacing the tuple in the indexes.

Let's separate concerns and have a common replace code path. We will then simply notify the transaction manager about individual index replaces, and it will take care of the read/write set tracking.

This way the transaction management API will be easy to scale for multikey indexes (#6385), which require a series of ordinary index replaces.