fix(vllm): support legacy and native request failure protocols

[ketor]

Compatibility correction

v2.27.1 required the native transfer-results API. This change supports legacy and native engines with the same connector, without editing installed vLLM source files.

• Restore the legacy get_finished + worker metadata path while retaining native get_transfer_results.
• Carry actual failed request IDs through legacy metadata, including nested MultiConnector routing, and gate recovery on aggregated receive completion.
• Adapt the old scheduler's existing recompute/fail funnel without fabricated block IDs or copied engine loops; preserve current foreign metadata and cancelled-request ownership.
• Keep serialized I/O fences and per-request failed-source quarantine. Native engines retain their native scheduler path.
• Add regression coverage through real scheduler/executor/MultiConnector machinery and register CPU lifecycle checks in CTest.

Verification so far

• Actual vLLM 0.29.0 and legacy DSV4.1 build 179dd0fa9: each 199 passed plus 24 subtests. Optional storage test and native-only branch are excluded on those runs.
• Native vLLM 8b1d188: 170 passed plus 24 subtests; legacy-only tests execute on the legacy builds instead.
• Standalone CPU connector/worker suites: 13 and 25 tests passed.
• Two independent static reviews; identified MultiConnector metadata loss was fixed and re-reviewed.
• Real-model dual-engine hardware fault qualification is in progress; PR remains draft until complete.

Support is capability-based within the V1 connector lifecycle; this does not claim arbitrary historical vLLM V0 engines or partial third-party protocol backports. Native ABI, storage keys and wire format are unchanged.