openai_legacy provider drops reasoning content, causing APIEmptyResponseError #1155

Description

When using kimi-cli with an OpenAI-compatible server (e.g. sglang or vllm) that separates reasoning/thinking content into a dedicated response field, the openai_legacy provider drops all reasoning content because reasoning_key is never passed to the OpenAILegacy constructor.

The underlying kosong library's OpenAILegacy class already supports a reasoning_key constructor parameter that tells the stream handler which field to read thinking content from. But create_llm() in llm.py never passes it:

case "openai_legacy":
    chat_provider = OpenAILegacy(
        model=model.model,
        base_url=provider.base_url,
        api_key=resolved_api_key,
    )

Impact

When the model produces a response consisting solely of reasoning/thinking (no text content, no tool calls), the stream yields zero parts and kosong/_generate.py raises APIEmptyResponseError("The API returned an empty response."). The agent crashes even though the server returned a valid response.

Proposed fix

1. Add an optional reasoning_key: str | None = None field to LLMProvider in config.py.
2. Pass reasoning_key=provider.reasoning_key when constructing OpenAILegacy in create_llm().

This allows users to configure e.g.:

[providers.vllm]
type = "openai_legacy"
base_url = "http://localhost:8000/v1"
api_key = "dummy"
reasoning_key = "reasoning_content"

PR: #1154