feat(cache): add ssd_write_drops counter for write-queue saturation

[ivaniguarans]

Summary

Adds a runtime counter ssd_write_drops to PagedSSDCacheManager that increments at every site where the background writer queue is saturated and a cache write is dropped or skipped. The new metric pairs with the existing read-path counters (hits, loads, hot_cache_hits, etc.) in PagedSSDCacheStats, giving operators a first write-path health signal. WARN logs at these sites already exist; this PR makes the drops aggregable so they can be graphed, alerted on, or surfaced through the runtime cache observability surface added in #1183.

Purely additive: no change to drop behaviour, no change to any cache code path, no new dependency.

Changes

• omlx/cache/stats.py: new ssd_write_drops: int = 0 field on PagedSSDCacheStats, grouped with the other operation counters (saves, loads, errors); added to reset() for parity with the other runtime counters.
• omlx/cache/paged_ssd_cache.py: "ssd_write_drops": 0 initialised in _stats; incremented at all three queue-saturation drop sites in this module — the hot-cache eviction path in _enqueue_ssd_write, the cold-store preflight _write_queue.full() guard, and the cold-store late except queue.Full fallback. Passed through in get_stats() and get_stats_for_model(). get_stats_dict() already spreads **self._stats, so no manual update needed there.
• tests/test_hot_cache.py: new TestSSDWriteDrops class with five tests — dataclass default + reset(), wiring round-trip through both stats accessors, and one test per drop site. The drop-site tests use unittest.mock.patch.object on the queue methods for deterministic firing rather than depending on writer-thread timing.

The bulk-eviction unlink path is intentionally excluded — that fallback runs an inline unlink() when the unlink queue saturates, but the file already exists on disk so no cache write is dropped. Counting it would conflate cleanup-queue backpressure with actual write loss.

Test plan

• uv run pytest tests/ -x — full suite passes
• uv run pytest tests/test_hot_cache.py::TestSSDWriteDrops -v — 5/5 pass
• Optional manual verification: load a small model (e.g. Qwen3-0.6B-4bit or Qwen3-Coder-Next-6bit per the project's cache micro-benchmarking convention), reduce _MAX_PENDING_WRITES to force queue saturation under a write-heavy workload, confirm ssd_write_drops appears and increments in /admin/api/stats.

[jundot]

Thanks for this. Verified the 3 drop sites cover real write loss, and the new field shows up in admin stats without extra wiring. Merging.