Send max_completion_tokens, not max_tokens, on the openai_compatible path

Send max_completion_tokens, not max_tokens, on the openai_compatible path Every current OpenAI model (the GPT-5 line) hard-rejects max_tokens with a 400 and requires max_completion_tokens; gpt-4o-mini and current OpenAI-compatible vendors accept the newer name. Sending both is impossible -- the presence of max_tokens alone triggers the 400 -- so the wire field is switched outright.