diag(cache): log which tier answered a block lookup, behind VMLX_CACHE_HASH_DEBUG

A FETCH line alone does not say whether L2 was consulted, and "block found
then discarded by a policy gate" is indistinguishable from "block not found"
from outside the process. Two lines close that: L1MISS (with whether a disk
store is attached) and L2READ -> HIT/miss.

They resolved a multimodal-reuse investigation that four separate controls had
failed to explain. The decisive observation was that on a media follow-up turn
the first miss is at pos=26, not pos=0 -- blocks 0-25 were FOUND (1,664 tokens)
and the turn still reported cached=0, because _mllm_request_has_media_cache_context
correctly refuses to reuse token-keyed KV for a prompt carrying media
placeholders. Not a lookup bug; the media-keyed restore leg is simply missing.

Off by default; zero cost when the flag is unset.