concurrency: improve lock table verification

In test builds, we run verification of the lock table on every update. This verification is useful and has helped discover bugs in the lock table.

However, we've recently encountered two problems we might want to address somehow:

1. Under deadlock builds, we've seen failures that _appear_ to be related to lock table verification being very slow.
2. A lock table verification failure can actually happen because of some _other_ request other than the one doing the verification because of how locking works.

For (1), one potential idea is to only run verification after a given proposal has done all of it's lock table updates. This would mean that a request that discovers or updates 10 locks would only have to run verification once rather than 10 times.

Jira issue: CRDB-59084