feat: capture request metadata

[Alex R (alexr17)]

fixes: #55

Summary

OpenAI and Anthropic instrumentation now captures a braintrust.metadata tag on LLM spans containing all request parameters except messages — including model, temperature, max_tokens, tools, tool_choice, top_p, and
more. This mirrors the Python SDK behavior where non-message params are separated into metadata for trace observability.

Changes

• OpenAI (baggage path): Extracts all non-messages fields from the raw request JSON into braintrust.metadata (captures the full wire request including tools, temperature, max_tokens, response_format, etc.)
• OpenAI (fallback path): Manually extracts temperature, max_tokens, top_p, frequency_penalty, presence_penalty, stop, end_user_id, tools, tool_choice, and response_format from ChatCompletionOptions
• Anthropic: Captures model, max_tokens, temperature, top_p, top_k, stop_sequences, tools, and tool_choice from MessageCreateParams in both streaming and non-streaming paths
• Tests: Added braintrust.metadata assertions to ChatCompletion_CapturesRequestAndResponse (both transport and fallback paths) and MessageCreation_CapturesRequestAndResponse

Format

  {
    "provider": "anthropic",
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 512,
    "temperature": 0.7
  }

[Andrew Kent (realark)]

LGTM -- if you haven't already, I'd recommend running the examples to make sure this shows up as expected in the UI (I'm pretty sure this is correct but usually good to verify)