Add Kimi K2 native tool calling for chat.json models

[c85199917-sudo]

Summary

Add native Kimi K2 tool-calling support for models served through
chat.json.

Kimi K2 tokenizers can carry native tool-call control tokens even when
chat.json itself only describes ordinary conversation turns. This change
discovers those tokens when available and uses them to support
OpenAI-compatible tool declarations, assistant tool calls, tool-result
continuations, and streaming tool-call responses.

Changes

• discover Kimi K2 native tool markers from the tokenizer
• render tool declarations in Kimi K2's native format
• render assistant tool-call and tool-result conversation turns
• parse generated Kimi K2 tool calls into OpenAI-compatible tool_calls
• support streaming tool-call deltas
• return finish_reason: "tool_calls" when appropriate
• preserve tool refusal for chat.json containers without the required
  native tool markers
• report conversation + native tools at server startup when the
  capability is detected
• support DeepSeek expert names in the container verifier

Validation

• full test suite: 228 tests passed, 2 skipped
• git diff --check clean
• tested against the real Kimi K2 WASTE container
• verified non-streaming native tool-call generation
• verified OpenAI-compatible tool_calls
• verified finish_reason: "tool_calls"
• verified tool-result continuation to a final assistant response

Example live response:

{
  "message": {
    "role": "assistant",
    "content": "I'll check the weather for Paris for you.",
    "tool_calls": [
      {
        "id": "call_0",
        "type": "function",
        "function": {
          "name": "get_weather",
          "arguments": "{\"city\": \"Paris\"}"
        }
      }
    ]
  },
  "finish_reason": "tool_calls"
}