Add vLLM NIXL PD smoke compatibility

[AAbouzeid]

Summary

• add an eager vLLM NixlConnector compatibility patch for kvcached PD disaggregation
• align NIXL registration with kvcached KV tensor block counts and keep kvcached NIXL on the NHD path
• add a hard guard for KVCACHED_CONTIGUOUS_LAYOUT=true with NIXL to avoid silent KV corruption
• add a GPU smoke script that compares plain vLLM+NIXL against kvcached+NIXL

Validation

• python -m pytest tests/test_vllm_nixl_compat.py -q
• bash -n tools/run_vllm_nixl_pd_smoke.sh
• pod smoke: INSTALL_VLLM=0 ./tools/run_vllm_nixl_pd_smoke.sh
• pod semantic smoke prompts: arithmetic retrieval, ticket lookup, checksum extraction; each passed with and without kvcached

[cui36]

/gemini review

[gemini-code-assist]

Code Review

This pull request implements monkey-patches for vLLM's NixlConnector to ensure compatibility with kvcached during PD disaggregation. The changes force an NHD layout and synchronize the physical block count in NixlConnectorWorker with kvcached's internal allocation. A review comment identifies that the current block count synchronization only handles over-allocation and suggests updating the logic to handle any mismatch to avoid assertion failures.

[AAbouzeid]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces patches for vLLM's NixlConnector to ensure compatibility with kvcached PD disaggregation. It addresses layout incompatibilities by forcing NHD and resolves block count assertion failures by dynamically updating the connector's block count using a new tracking variable. Review feedback suggests that using a module-level global variable for this state is brittle and may cause issues in multi-engine or multi-tenant scenarios, recommending a more explicit configuration management approach instead.