test: Verify and document the first OpenDAL cache engine

Part of #1350. Verify the concrete first-version behavior of #1357 / #1360 and the example in #1361 as they are implemented. This is the delivery evidence checklist, not a separate exhaustive acceptance project.

First-version blockers

• Exercise the actual public construction path on Fs and real Redis: memory hit, secondary hit after memory eviction, source fetch, supported write policies, duplicate submissions, version isolation, eviction/refill and close.
• Preserve focused deterministic coverage for the implemented publication/registration rules, including out-of-order duplicate submissions and old work confined to its immutable version. Strict mutable-key invalidation is excluded.
• Verify the selected resource and failure contract: full/oversized admission, representative slow or failing IO, budget release, read fallback, bounded shutdown and best-effort cleanup. Test the documented no-recovery/namespace behavior rather than unsupported recovery or sharing.
• Run relevant feature/build compatibility checks and existing BlockEngine regressions for changed shared paths. Keep documented examples executable.
• Document setup, immutable keys and exact-version source reads, limit/metric meanings, failure behavior and ownership of residual objects. Link implementation revisions and reproducible evidence for the required checks.

Reuse valid existing validation and add coverage for changed paths or concrete gaps. Complete this alongside the engine; possible follow-ups do not keep this issue open. Builds and experiments continue on remote infrastructure rather than the requester's local machine.

Possible follow-ups — not blockers

A full latency/throughput/resource matrix across workloads and backends, long-duration stress campaigns, comprehensive fault injection and additional backend certification. No performance improvement is required or implied by introducing the engine. Any performance claim still needs a relevant same-environment measurement; prior prototype results must retain their original attribution.

The first-version checks establish only the documented Fs/Redis scope, not S3, multi-machine sharing or crash recovery.