fix: cap CUDA IPC multimodal pool budget

[mickqian]

Summary

• treat SGLANG_MM_FEATURE_CACHE_MB as a hard per-node CUDA IPC pool budget across tokenizer workers
• report the actual total and per-worker HBM reservation at startup
• add CPU-only regression coverage for budget splitting

Root cause

Each tokenizer worker owns a distinct allocation on base_gpu_id. The previous 128 MiB per-worker floor could make the aggregate reservation larger than the configured budget (for example, a 1 GiB budget with 16 workers reserved 2 GiB).

Behavior

The pool now splits the configured budget exactly across workers. A pool miss retains the existing CPU-transport fallback, so SGLang does not allocate additional CUDA memory to preserve IPC.

Validation

• PYTHONPATH=$PWD/python uv run python -m pytest test/registered/unit/multimodal/test_cuda_ipc_pool_budget.py -v
• direct-file unittest invocation
• pre-commit hooks (including registered-test CI validation)

---

CI States

Latest PR Test (Base): ❌ Run #29174721443
Latest PR Test (Extra): ❌ Run #29174721370

[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]

closing due to duplication with #31227