fix(vllm): recover hybrid KV loads through request-level outcomes

[ketor]

Summary

• Release candidate v2.27.1: migrate hybrid/multi-group loads to native request-level completion/failure outcomes; never emit ambiguous block errors for those layouts.
• Separate scheduler waiting from I/O execution. Serialized loads retain explicit owner-device pre/post fences, including runners that defer the load hook until after forward.
• Enforce the request-level completion fence even with the legacy DFKV_GPU_LOAD_FENCE=0 override.
• Preserve cancellation/ownership and all-rank completion, prevent repeated submission, and quarantine failed external hits until request completion.
• Remove the obsolete scheduler invalid-block patch. Native ABI, wire format, storage keys and functional C++ code are unchanged.

Verification

• Pinned real vLLM runtime: 8b1d188046034b42024d22ee52fa6c3948b4acbd ([Bugfix][Mooncake] Report request-level KV load failures under HMA vllm-project/vllm#56855).
• Connector suite: 158 passed plus 24 subtests; optional live-storage case independently passed with actual GPUDirect RDMA GPU byte comparison.
• Real multi-rank model qualification: async/serialized I/O × recompute/fail policy; healthy external loads, persistent genuine store NotFound after lookup, injected native-completion exception, delayed peer completion, correct subsequent requests.
• Cancellation during a pending load retained ownership until completion; subsequent request and prefix-cache reset succeeded.
• Unrelated cold prefill remained usable during a delayed failed load; asynchronous execution made independent progress.
• Real dual-device completion regression includes the legacy fence opt-out; pre-fix test failed and corrected contract passed.
• Three v2.27.1 wheels built and imported in the pinned engine environment; release-contract tests passed.
• Two independent static reviews plus final inline fence review. All eight checks on final head 02bb44d are green.

Compatibility

Requires the complete native get_transfer_results / failed_recving protocol in the vLLM runner, executor and scheduler. Upgrade older engines lacking that API together with this connector; no compatibility shim or scheduler patch is shipped.