feat: unify multimodal feature transport

[mickqian]

Summary

• add --mm-feature-transport={cpu,cuda_ipc} as the single NVIDIA CUDA IPC policy
• map --keep-mm-feature-on-device and SGLANG_USE_CUDA_IPC_TRANSPORT to the new policy with deprecation warnings
• reject CUDA IPC on non-NVIDIA and multi-node configurations
• log the fixed pool reservation, base GPU, tokenizer-worker count, KV-cache headroom effect, and CPU fallback

Memory policy

CUDA IPC remains opt-in (cpu is the default). On H200 TP8 with Kimi-K2.7-Code, the existing 1 GiB pool is allocated immediately on GPU 0; Kimi's largest raw image feature is about 36.75 MiB, so that pool holds about 27 worst-case images. This fixed reservation reduces KV-cache headroom and should not become an unconditional default.

Dependency

This PR depends on #30902, which makes SGLANG_MM_FEATURE_CACHE_MB a hard aggregate budget across tokenizer workers. It should merge first.

Validation

• py_compile for modified Python files
• pre-commit hooks, including formatter and registered-test CI validation
• focused unit-test collection is blocked locally before collection by an existing huggingface_hub==1.6.0 / @strict incompatibility; CI will run the tests in the pinned environment.

---

CI States

Latest PR Test (Base): ❌ Run #29379619765
Latest PR Test (Extra): ❌ Run #29379619656

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[mickqian]

/tag-and-rerun-ci

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!