Improve locking semantics for UPDATE / UPSERT / DELETE

When the master node receives an `UPDATE`, and has default config settings, it acquires an exclusive lock on the shard id. This leads to a notable performance degradation for `UPDATE`, `INSERT ... ON CONFLICT UPDATE`, and `DELETE` statements.

If the user enables `all_modifications_commutative`, we instead acquire a shared lock.

This issue relates to investigating and implementing a locking policy that improves performance for `UPDATE` statements.

No files or tests are named. Start by tracing how the master node locks the shard id for UPDATE, INSERT ... ON CONFLICT UPDATE, and DELETE, then compare the default behavior with all_modifications_commutative. Done means a defined locking policy is implemented and its correctness and performance are validated.

Written by the indexing model from the issue text.