Fixing Tool-Call Accuracy in vLLM

A vLLM fork achieved 100% tool call schema accuracy with Kimi and DeepSeek models, and vLLM 0.31's constrained decoding support now enables the same result with newer versions of those models.