test(vllm): verify preemption lifetimes and release 2.26.3

[ketor]

Changes

• Replace forwarding-only preemption coverage with a source-byte lifetime regression using actual connector/worker/sender and the installed engine pre-forward hook when available.
• Cover consumer MRV1 synchronous/asynchronous resume, prior/same-step preemption, continued prefill, decode and completion.
• Document fresh model_revision guidance for potentially corrupted historical objects; format compatibility does not certify content integrity.
• Synchronize release metadata to 2.26.3. No connector runtime or C++ functional changes since 2.26.2.

Executed acceptance

• Actual target runtime: 119 tests passed, 24 subtests passed, one DFKV_MEMBERS-dependent test skipped.
• Behavioral A/B: three expected assertions fail on 2.26.1; all 15 selected cases pass on 2.26.2.
• Separate CUDA/native RDMA probe: four rounds, 12 PUTs/20 GETs, exact payload hashes, stale generation cancellation and destination reuse checks passed. Delays precede native calls; this is not a claim of observing NIC DMA mid-flight.
• Isolated real-model execution: three forced preemptions, validated output, positive native GET/PUT and zero completion-error delta.
• Cold and fresh-process restart quality: 9/9 each; positive external-cache token hits after restart.
• Local three release wheels built; Ruff test checks passed.

Tests in this PR ran in the actual target runtime. Generic hosted CI alone does not constitute vLLM/CUDA/RDMA acceptance.