fix: use max_tokens for openai provider with custom base URL (Mistral compatibility)

[octo-patch]

Fixes #852

Problem

When using HINDSIGHT_API_LLM_PROVIDER=openai with HINDSIGHT_API_LLM_BASE_URL pointing to Mistral's API (https://api.mistral.ai/v1), Hindsight crashes on startup during verify_connection because max_completion_tokens is sent but Mistral only accepts max_tokens, returning a 422:

openai.UnprocessableEntityError: Error code: 422 - {'message': {'detail': [{'type': 'extra_forbidden', 'loc': ['body', 'max_completion_tokens'], ...}]}}
RuntimeError: Connection verification failed for openai/mistral-small-latest

Solution

Add _max_tokens_param_name() method that selects the correct parameter name based on provider and base URL:

• Native OpenAI (no custom base_url) → max_completion_tokens
• Groq → max_completion_tokens (Groq supports the newer name)
• openai with custom base_url, ollama, lmstudio, minimax, volcano → max_tokens

A custom base_url on the openai provider signals a third-party compatible API (Mistral, Together AI, etc.) that may not have adopted max_completion_tokens.

Testing

• Existing tests pass unchanged (they test the call() argument name, not the API key)
• The fix covers both call() and call_with_tools()