test(persistence): snapshot crash-scenario fail-points

[petrpan26]

Depends on the fail-parallel bootstrap (sibling chore(test) issue under area: persistence). Sibling to the WAL crash-coverage issue.

Cover the snapshot write / read crash surface.

Fail-points to add:

• snapshot::write::post_header_pre_body — header on disk, body interrupted
• snapshot::write::post_body_pre_rename — partial .tmp exists, atomic rename never happened
• snapshot::read::pre_crc_check — body read but CRC not yet verified
• snapshot::read::truncated_body — short read mid-stream
• recovery::pick_snapshot_vs_wal_tail — both present, simulate ambiguous state

For each, an integration test:

1. Build snapshot state via real pushes.
2. Trigger the fail-point.
3. Restart and assert recovery falls back correctly — older snapshot + WAL replay, or rebuild from WAL only if the latest snapshot is unusable.
4. Verify no .tmp files leak across restart.

Pair with the SNAPSHOT-UPGRADE deferral whenever a real FORMAT_VERSION bump arrives; this issue is the crash-correctness scaffolding it would build on. ~8 tests, ~400 LOC.