Detect model family for `max_completion_tokens` vs `max_tokens` (covers self-hosted OpenAI-compatible backends)

[ameliakuang]

After #3 we always send max_completion_tokens. We can add model-prefix detection for OpenAI-compatible endpoints: pick max_completion_tokens for gpt-4o, gpt-4.1, gpt-5, o1, o3, or o4, else max_tokens.

References:
NousResearch/hermes-agent#15377