tests: isolate failpoint lock counts into per-test key bands

The txn_cleanup_* failpoint tests share one cluster and run serially, and they
counted locks with a whole-keyspace scan. Async-commit locks resolve in the
background and drain at an unpredictable rate, so a later test's unbounded scan
would catch an earlier test's still-draining residue and over-count (flaky
'575 vs 512'; #525). txn_cleanup_locks_batch_size is the worst offender: it stubs
resolution via before-cleanup-locks and leaves 512 async-commit locks behind.

Fix, test-only. Give each lock-counting case its own disjoint band of the u32 key
space and scope both its lock scans and its cleanup_locks calls to that band, so a
neighbour's residual locks are never fetched or counted. Single-key tests scan only
their own key. The bands are a large fraction of the key space rather than a
one-byte prefix, so under MULTI_REGION each still spans several regions and the
cross-region cleanup/retry paths these tests exist to exercise stay covered.

- Partition the key space into NUM_KEY_BANDS bands; write_data generates keys inside
  a given band; count_locks_in_band / wait_for_locks_count_in_band scope the scans;
  count_locks_of_key scopes the single-key tests.
- Scope every cleanup_locks call to the acting test's band (the batch-size test's
  resolved_locks assertion depended on this).
- Drop the whole-keyspace helpers.

Suite is stable across repeated runs locally, including MULTI_REGION=1; no client
(src/) change.

Signed-off-by: Eduard R. <eduard@ralphovi.net>

Author: eduralph