refactor(kvcache): fold prefix matching into one accumulator

[nilig]

What type of PR is this?
/kind cleanup

What this PR does / why we need it:
Longest-prefix matching is implemented three times: LongestPrefixScorer over a materialized Lookup for the weighted score, and the precise-prefix producer's matchedBlockCount and matchedBlockCountByTier for the block and per-tier counts it publishes, re-walking the lookup result once per endpoint. No single place defines the contract, and the re-walk grows with fleet size.

Indexer.MatchBlockKeys becomes the one entry point. A pooled accumulator implements the rules once (candidates are the pods holding the first key, a chain ends at the pod's first missing key, duplicate entries at a key take the highest weight, every tier tracks its own chain) and folds each key's entries in one call. ScoreTokens and LongestPrefixScorer.Score are projections of it; the producer only aggregates per prompt and publishes, all or nothing. KVBlockScorer and NewTracedScorer stay as deprecated compatibility surfaces; existing score_tokens span attributes keep their meaning, and the matcher's own match_block_keys span reports the longest chain.

Ownership after this change: kvblock owns storage and reads; kvcache owns the one matcher, its fallback, and match telemetry; the producer owns prompt aggregation and publication. Index reads are unchanged: this PR materializes Lookup. A follow-up adds an ordered index walk that feeds the same accumulator without materializing.

M4 Max, n=6, ScoreTokens on 1,034 blocks held by 40 endpoints with eight rank entries each (BenchmarkScoreTokensKermit): 294.5ms and 4.6 GiB per request on main, 67.7ms and 116.5 MiB here.

Which issue(s) this PR fixes:
Part of #2390

Release note:

The precise prefix-cache producer and `kvcache.Indexer.ScoreTokens` share one prefix matcher (`Indexer.MatchBlockKeys`), which emits a `match_block_keys` span. `ScoreTokens` no longer emits `compute_scores`; `kvcache.KVBlockScorer` and `kvcache.NewTracedScorer` are deprecated.

Test plan:

• matcher parity with the replaced scorer and producer counters on 500 random fixtures (filters, gaps, duplicate pod and rank entries, unconfigured tiers, speculative entries)
• producer aggregation across prompts, multimodal match, candidate filtering, error and cancellation publish nothing
• score_tokens telemetry keeps its meaning (span test)
• make presubmit