fix(llm): add GLM compatibility for reasoning_content field

[jknightai]

Summary

Add compatibility for Zhipu GLM responses where text may appear in reasoning_content when content is empty.

Changes

• packages/core/src/llm/provider.ts
• Streaming path: accept delta.reasoning_content as output fallback.
• Sync path: fallback to message.reasoning_content when message.content is empty.
• Uses type assertions to keep TypeScript build passing.

Why

Some GLM models can return:

• content: ""
• reasoning_content: "actual text"

Without this fallback, InkOS may throw LLM returned empty response.

Validation

• pnpm -r build passes after changes.
• Runtime behavior verified with GLM endpoint.

[jknightai]

我这边补充下本地验证结果：在 GLM 上能复现 content 为空、实际文本在 reasoning_content 的情况，所以修复前会报 LLM returned empty response。这次我把同步和流式两条路径都做了 fallback，并用类型断言处理了非标准字段；标准 OpenAI 的 content 返回路径不受影响。
I’d like to add my local verification result: on GLM, I can reproduce cases where content is empty while the actual text is in reasoning_content, which caused LLM returned empty response before this fix. In this PR, I added fallback handling in both sync and streaming paths, and used type assertions for the non-standard fields; standard OpenAI-style content behavior is unchanged.

[Narcooo]

感谢 PR！审过代码，改动精准，流式和同步两条路径都覆盖了，不影响标准 OpenAI/Anthropic 通路。

预计将合入。

Thanks for the PR! Reviewed the code — clean, precise change that covers both streaming and sync paths without affecting standard OpenAI/Anthropic flow.

Will merge.