fix(providers): add max_completion_tokens for openai o1 compatibility

[flobo3]

Fixes #2462

What
Adds max_completion_tokens alongside max_tokens in the openai_compat_provider payload.

Why
OpenAI deprecated max_tokens in favor of max_completion_tokens starting with the o1 model series to separate the limit for final output tokens from the new "reasoning tokens". When using o1 models via OpenAI-compatible endpoints, passing only max_tokens results in an unsupported_parameter error.

How it works
We now send both max_tokens and max_completion_tokens with the same value. This ensures compatibility with newer endpoints that require max_completion_tokens (like OpenAI's o1 models or recent vLLM/Ollama versions) while maintaining backward compatibility with older local servers that might strictly validate the payload and reject unknown parameters.

[chengyongru]

Thanks for your work! Could you make a PR 2 main branch?