Isolate the storage cache by storage namespace

[s-prosvirnin]

Problem

CacheKey holds nothing but the relative path. The key carries
no storage identity, while the cache state behind it is shared:

• StorageCache (foyer, the bytes) — one per process, shared by the Iceberg path
  and the object-store path;
• CacheLayer (stat cache, KeyLocks) — one per OperatorRegistry, cloned into
  every operator it builds;
• PrefetchLayer: seen: DashSet<Arc<str>> and
  InFlightTracker.pending: DashMap<Arc<str>, _>— keyed by the
  bare path as well.

With two or more buckets (warehouse + catalog is the standard deployment) this
means:

• a read of tables/logs/metadata/v1.metadata.json in the second bucket can be
  served the first bucket's bytes;
• a stat (HEAD) against the second bucket returns the first bucket's metadata;
• a delete or write in one bucket invalidates the other bucket's stat entry;
• prefetch treats an object as already seen because of a different bucket and
  skips the read-ahead.

Risk: wrong data in query results, and GC deleting objects it did not mean to.

Fix

1. Introduce StorageNamespace — the address of the data, credentials excluded:
   endpoint, bucket, root. Compare it structurally; if a compact key is
   wanted, derive a collision-resistant digest, not a 64-bit DefaultHasher.
2. Put the namespace into CacheKey and into the keys of stat_cache,
   KeyLocks, the prefetch seen set and InFlightTracker.
3. Thread the namespace from registry::OperatorKey to the point where
   CacheLayer and PrefetchLayer are applied. StorageLayers::wrap_operator
   currently does not know which operator it wraps, so its signature changes.

Alternatives rejected:

• a separate CacheLayer per operator — does not fix it, StorageCache stays
  shared;
• a separate foyer cache per bucket — splits the shared memory budget and
  duplicates the disk cache.

Compatibility

Changing CacheKey changes the serialized key, i.e. foyer's on-disk format.
Existing disk-cache entries become misses. It warms up after the deploy; no
migration needed.

Definition of done

• a cache hit never crosses endpoint/bucket/root;
• the same relative path in two buckets returns the expected, different bytes and
  different metadata;
• a delete or write in one bucket does not invalidate the other's stat entry;
• rotating credentials under an unchanged namespace keeps the cache shared —
  otherwise the operator-leak fix loses its point for a credential-vending
  catalog;
• the test proves an actual cache hit (via CacheMetrics counters), not merely a
  successful read from the backend;
• coverage: a component test for the path collision plus an S3 integration test
  over two buckets.