fix(llm_gateway): make token breakdown fields optional and support float penalties

[pylrn]

Summary

When LLM Gateway proxies models that return partial token details (such as OpenAI reasoning models returning completion_tokens_details: {"reasoning_tokens": 128} without audio/prediction fields, or Claude models returning prompt_tokens_details: {"cached_tokens": 512} without audio fields), Pydantic raised a ValidationError because all breakdown fields were declared as required integers.

This PR makes token detail breakdown fields optional with a default of None, adds thinking delta support for streamed reasoning models, and ensures model penalty parameters support float values.

Changes

• assemblyai/llm_gateway/v1/models.py:
  • In LLMGatewayCacheCreation, LLMGatewayPromptTokensDetails, and LLMGatewayCompletionTokensDetails, declare breakdown fields as Optional[int] = None so partial breakdowns from upstream providers parse cleanly.
  • In LLMGatewayChunkDelta, add thinking: Optional[str] = None for parity with LLMGatewayResponseMessage.thinking during streamed completions with reasoning/thinking models.
  • In LLMGatewayDefaultParameters, update frequency_penalty to Optional[float] = None and add presence_penalty: Optional[float] = None.
• tests/unit/test_llm_gateway.py:
  • Add test_chat_completions_tolerates_partial_token_details covering partial prompt caching and reasoning token breakdowns.
  • Add test_stream_chunk_exposes_thinking_delta verifying streamed reasoning delta parsing.
  • Add test_list_models_tolerates_float_penalties_and_presence_penalty verifying float model parameter parsing.
• assemblyai/__version__.py:
  • Bump to 1.6.1 to satisfy CI version bump check.

Backwards Compatibility

All new/modified fields are Optional with None defaults or widen integer types to floats, preserving full compatibility with existing code and responses.