fix: switch kimi-k2 to openai_json_xml tool dialect

The kimi_k2 vLLM section-token dialect stopped working for kimi-k2
and kimi-k2-thinking — Cascade returns ~16 token empty responses
instead of tool calls. Switching to openai_json_xml which works
for all Moonshot models.

Verified on VPS: kimi-k2 with openai_json_xml correctly emits
<tool_call> XML tool calls.