[stable34] fix(filecache): announce every removed entry so metadata is cleaned up

[backportbot]

Backport of #63998

Warning, This backport's changes differ from the original and might be incomplete ⚠️

Todo

• Review and resolve any conflicts
• Review and verify the backported changes
• Amend HEAD commit to remove the line stating to skip CI

---

Learn more about backports at https://docs.nextcloud.com/server/stable/go.php?to=developer-backports.

[f0rthelulz]

A production data point from 34.0.2 (MariaDB, Redis file locking), in support of getting this into 34.x. 34.0.2 still ships the unfixed Cache::removeChildren.

What happened. A folder subtree of about 462,000 cache entries had been removed from disk, and occ files:scan --shallow of its parent was run: one Cache::remove() into removeChildren(). It ran about 100 minutes inside a single transaction and did not finish.

• The processlist showed a continuous stream of DELETE FROM oc_files_metadata[_index] WHERE file_id IN (0, 1, 2, 3, …) at about 2 statements/s.
• Handler_delete did not move, so each statement deleted 0 rows. The ids are array_chunk positions, because keys are not preserved, and every chunk re-sends the accumulated $cacheEntryRemovedEvents.
• By the arithmetic of the unfixed loop, 462 chunks means about Σ2k ≈ 214,000 metadata DELETEs with IN lists growing towards 462k values, plus about 107M per-entry CacheEntryRemovedEvent dispatches. That is days of work.
• The open transaction held row locks the whole time, so other writes, cron included, failed with 1205 Lock wait timeout exceeded.
• A background cron job (Text's Cleanup, going through the change watcher) independently started the same removal on the same missing tree and held the locks too.

The workaround. SIGKILL rolled both back cleanly. We then recreated the empty folder skeleton, let a recursive files:scan drop the file rows one at a time, and removed the folders leaves-first, so that no removal ever had cached descendants. The two lines in this PR (a fresh $cacheEntryRemovedEvents per chunk, and array_chunk(…, true)) are exactly what we had tried to patch in locally.

The metadata half, drop*ForFiles binding $fileIds instead of $chunk, is also in 34.0.2, but it is already backported (#62771, #64092).