feat(usage): per-model and per-day prefix-cache hit-rate observability

[OnlyTerp]

Summary

Surface prefix-cache hit-rate in the usage dashboard, per model and per day. The aggregate already records cache_read/cache_write tokens; this exposes them in the breakdowns and UI.

Why (local-model)

Lets you see whether local prefix caching is actually helping — e.g. after the tool-order normalization PR — and how the hit rate trends over time. Inference-efficiency signal, not API cost.

Changes

• controller/src/stores/inference-request-store.ts: add cache_read/cache_write/cache_hit_rate to the by_model, daily, and daily_by_model breakdowns; extract a shared calcCacheHitRate (also used by the global hit-rate).
• controller/src/modules/system/usage/pi-sessions.ts: per-model / per-day cache accumulation in the parallel Pi-sessions aggregator.
• shared/contracts/usage.ts + frontend normalizer: additive cache fields (default 0).
• UI: a sortable "Cache hit %" column in the per-model table, plus a per-day readout and an "avg cache hit" summary stat in the daily chart.

Verification

• cd controller && bun test src/stores/inference-request-store.test.ts → pass; bun run typecheck → clean.
• frontend typecheck + lint + build → clean.

Non-goals

• Additive only; no contract removals. Frontend normalizer defaults missing fields to 0 for older payloads.