feat: Support reasoning_content in OpenAI chat completions streaming response

[tc3oliver]

Summary

Add support for reasoning_content field in streaming responses from OpenAI-compatible APIs.

Problem

Some LLM inference engines (e.g., vLLM with reasoning models like DeepSeek-R1 or QwQ)
return streaming content in the reasoning_content field instead of content.
This causes the benchmark to incorrectly report output_tokens = 1 because
the actual generated text is not captured.

Solution

Check both content and reasoning_content fields when processing streaming chunks.
This maintains backward compatibility while adding support for reasoning models.

Testing

Tested with:

• vLLM serving a 120B reasoning model (uses reasoning_content)
• Ollama serving llama3:70b (uses standard content)

Both scenarios now correctly capture output tokens and metrics.