fix(vlm): return invalid media input as HTTP 400

[mickqian]

Motivation

Invalid client-supplied multimodal data (for example malformed base64 image bytes) is classified as ValueError by _load_single_item, but both fast and legacy loaders wrap it in RuntimeError. The OpenAI serving layer therefore returns HTTP 500 instead of HTTP 400.

Changes

• Preserve ValueError through both multimodal loading paths.
• Keep unexpected loader failures as RuntimeError/HTTP 500.
• Add a CPU regression test for the fast loader.

Validation

• test_base_processor_image_decode.py, Kimi processor tests, and OpenAI serving-chat unit tests: 101 passed + 15 subtests on NVIDIA B300.
• Real Kimi-K3 TP8/GB300 protocol probe before this patch reproduced malformed base64 image as HTTP 500; post-patch validation at Kimi PR head bac10440f on real Kimi-K3 weights, NVIDIA GB300 TP8, returned HTTP 400. The complete protocol/error matrix passed 5/5, and the server log contained 0 Traceback, 0 ERROR.

---

CI States

Latest PR Test (Base): ❌ Run #29830822824
Latest PR Test (Extra): ❌ Run #29830822508

[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