Use finer-grained locks to enhance reference table write throughput

We currently use shard-level locks to serialise writes to replicated/reference tables, only allowing inserts to run concurrently. In the case of a reference table, shard-level is the same as table-level. This significantly limits write concurrency and thus overall throughput.

For updates/deletes/inserts that are known not to conflict because they have disjoint primary key filters, we could be using more fine-grained locks on primary key values or ranges of primary keys. A thing to watch out for is updates to the primary key value (should perhaps still use a global lock).

No files, tests, or entry points are named. Start by tracing shard-level locking for replicated/reference table writes, then evaluate primary-key or primary-key-range locks for disjoint inserts, updates, and deletes, including the proposed global-lock behavior for primary-key updates.

Written by the indexing model from the issue text.