feat(types): add presence_penalty and token breakdown details to LLM Gateway

[pylrn]

Description

This PR brings the Node.js SDK's LLM Gateway types into parity with the AssemblyAI LLM Gateway API and the Python SDK (parity with Python SDK PR #261).

Changes:

• ModelDefaultParameters: Expose presence_penalty?: number | null;
• ChatCompletionRequest: Expose presence_penalty?: number; and frequency_penalty?: number;
• ChatCompletionMessage: Expose thinking?: string | null; to support reasoning models returning thought processes
• ChatCompletionUsage: Expose detailed token breakdown fields:
  • prompt_tokens_details?: ChatCompletionPromptTokensDetails | null; (including cached_tokens, audio_tokens, cache_creation with ephemeral_5m_input_tokens / ephemeral_1h_input_tokens, and cache_write_tokens)
  • completion_tokens_details?: ChatCompletionTokensDetails | null; (including reasoning_tokens, audio_tokens, accepted_prediction_tokens, rejected_prediction_tokens)
  • input_tokens?: number | null; and output_tokens?: number | null;
• ChatCompletionResponse: Expose optional http_status_code, response_time, llm_status_code, and request
• ChatCompletionChoice: Allow finish_reason to be null
• Added comprehensive unit tests in tests/unit/llm-gateway.test.ts verifying parsing of detailed usage breakdowns, thinking tokens, and presence penalties
• Updated CHANGELOG.md and bumped package.json to 4.41.6 for the Version Bump Check workflow