feat: add kwargs support for completion calls

[lakshyaag]

Summary

This PR adds support for passing additional kwargs through the entire RLM stack to underlying LM client APIs, enabling fine-grained control over completion parameters like max_tokens, temperature, reasoning_effort, and provider-specific options.

Key Changes

1. Enhanced Client Interface (rlm/clients/)

• Updated BaseLM abstract methods to accept model and **kwargs parameters
• Modified OpenAIClient.completion() and OpenAIClient.acompletion() to accept and forward kwargs to OpenAI API
• Improved _track_cost() to safely handle missing usage data with getattr() fallbacks

2. Core Communication Updates (rlm/core/)

• Extended LMRequest dataclass with optional kwargs field for passing parameters through socket/HTTP protocol
• Updated send_lm_request_batched() to accept and forward kwargs
• Modified LMHandler and LMRequestHandler to unpack and pass kwargs to client completion calls
• Updated RLM.completion() to accept kwargs and propagate them through:
  • _run_iteration() for main completions
  • _default_answer() for fallback completions
  • _fallback_answer() for max-depth fallback

3. Environment Integration (rlm/environments/)

• LocalREPL: Updated _llm_query() and _llm_query_batched() to accept and forward kwargs via socket
• ModalREPL: Updated broker exec script's llm_query() and llm_query_batched() to include kwargs in HTTP payloads
• DockerREPL: Updated proxy handler and exec script to pass kwargs through HTTP requests

4. Testing & Examples

• Added tests/clients/_openai.py with basic tests demonstrating:
  • Completion with kwargs (e.g., reasoning_effort="high")
  • Completion without kwargs (backward compatibility)
• Updated MockLM implementations in tests and examples to match new BaseLM signature

Backward Compatibility

All changes are fully backward compatible:

• **kwargs parameters default to empty, so existing code works unchanged
• model parameter formalizes existing pattern used by all client implementations
• No breaking changes to public APIs
• All existing tests should continue to pass

Example Usage

# Root-level RLM call with kwargs
rlm = RLM(backend="openai", model="gpt-4")
result = rlm.completion(
    prompt="Solve this problem", 
    max_tokens=1000, 
    temperature=0.7
)

Open questions

1. Should the sub-LLM be extended to inherit kwargs from the root LM? Another direction here is to let the root LM decide additional kwargs.