fix(ssd-cache): preserve original bfloat16 dtype across quantized spill

[CBribiescas]

Summary

Follow-up to #605 per @waybarrios's review comment: the dequant-on-spill path casts bf16 → fp16 (to dodge numpy's PEP 3118 buffer-protocol mismatch) but didn't record the original dtype. Reload then hands the model fp16 KV where it had computed bf16 — silent precision regression on every SSD reload of a quantized-cache model.

Root cause

_mx_to_numpy_safe only stamps a dtype hint when it did the upcast (bf16 → fp32 fallback). The quantized-spill path pre-casts bf16 → fp16 before snapshot_layer is called, and since fp16 IS numpy-supported, _mx_to_numpy_safe returns dtype=None. The snapshot then carries no signal that the data originated as bf16, and scheduler._reconstruct_ssd_layers has nothing to cast back from.

Fix

Mirror the existing plain-cache pattern: stash an explicit dtype sentinel on the layer before the cast, and have KVCacheSerializer.snapshot_layer honor it with priority over the autodetect path:

# enqueue_spill (dequant branch):
if str(getattr(k, "dtype", "")).endswith("bfloat16"):
    layer._ssd_keys_original_dtype = "bfloat16"
    layer._ssd_values_original_dtype = "bfloat16"
    layer.keys = k.astype(mx.float16)
    layer.values = v.astype(mx.float16)

# KVCacheSerializer.snapshot_layer:
keys_orig_dtype = (
    getattr(layer, "_ssd_keys_original_dtype", None) or keys_orig_dtype
)

Sentinel name is _ssd_-prefixed to avoid colliding with any mlx-lm-side cache attribute.

Also add "QuantizedKVCache": "supported_via_dequant_on_spill" to SERIALIZER_SUPPORT_MATRIX so the diagnostics table matches what enqueue_spill actually handles.

Reload side

No change. scheduler.py:3019-3028 already reads keys_original_dtype from the deserialized dict and casts back via mx.array(...).astype(mx.bfloat16). Only the spill-side signal was missing.

Tests

Four regression tests in TestLayerSerializer:

• test_support_matrix_includes_quantized_kv_cache — guards the matrix entry against drift.
• test_snapshot_layer_honors_original_dtype_sentinel — sentinel set, no autodetect signal: must record "bfloat16".
• test_snapshot_layer_sentinel_overrides_autodetect — sentinel wins even when _mx_to_numpy_safe also produces a hint.
• test_snapshot_layer_no_sentinel_falls_back_to_autodetect — plain fp16 layers still record no dtype hint (no regression on the existing path).

Verification

pytest tests/test_ssd_cache.py
63 passed in 3.09s

black --check vllm_mlx/ssd_cache.py tests/test_ssd_cache.py
2 files would be left unchanged

Refs #605 #563.