[Enhancement][Cache] Coalesce concurrent cache misses and support bounded stale refresh #7932

Goal

Reduce duplicate upstream work when many cache-eligible requests miss or expire on the same key, then optionally serve a bounded stale value while one refresh runs.

Existing capabilities and boundary

Extend the existing CachePlugin and reuse its cache eligibility, key construction, response/header and storage semantics. #7839 concerns Admin HTTP configuration-sync cache rebuilds, not business-response caching, and is a separate optimization.

MVP and staged extension

1. Add opt-in per-gateway single-flight for eligible requests sharing the effective cache key. One bounded upstream fill can supply a reusable cached representation to a bounded number of waiters.
2. Add stale-while-revalidate only after the miss path is correct, with distinct freshness/stale deadlines and a finite refresh budget. Preserve the old value only within the configured stale policy.

Start with finite, bounded HTTP responses. Do not share a one-shot response publisher or a DataBuffer across subscribers. Retain all isolation dimensions required by existing cache/auth/tenant/Vary policy; single-flight must not broaden cache eligibility or merge different callers' private responses.

Acceptance criteria

• A concurrent burst for one eligible cold key performs one fill within a gateway; distinct keys and security scopes stay independent.
• Bound active fill keys, waiters per key, response size, fill time and refresh concurrency. Overload behavior is documented and cannot create an unbounded second queue.
• Define leader/waiter cancellation, upstream errors, non-cacheable responses and eviction/invalidation during a fill. Failures release state and permit a subsequent fill.
• Invalidation/configuration generations prevent an old fill from repopulating a newly invalidated entry.
• Stale serving never exceeds its configured window; refresh failure, no stale value and concurrent invalidation are covered. No background retry storm.
• Preserve status, headers, body bytes and reference-counted buffer ownership; test with the real reactive response path.
• Document per-gateway scope: replicas can still perform separate fills. Distributed locking is not required for the first version.
• Compare upstream call counts, gateway latency and retained memory under cold-key and hot-expiry bursts using [Task][Performance] Establish reproducible benchmarks and track gateway optimizations #7650; disabled mode remains compatible.

Related: #7839 (different cache layer), #7650.