fix(openai): non-compatible providers send correct max_tokens field

[peteski22]

Summary

• DeepSeek and Mistral delegate to CompatibleProvider.Completion() which maps MaxTokens → max_completion_tokens on the wire. Both APIs actually expect max_tokens, meaning the token limit was silently dropped.
• Adds ChatCompletionRequestTransform hook to CompatibleConfig (Strategy pattern) so providers can adjust the SDK request after convertParams builds it.
• DeepSeek and Mistral supply transforms that swap max_completion_tokens → max_tokens and clear unsupported fields.
• Moves Mistral's user/reasoning_effort stripping from preprocessParams to transformRequest where it operates on the actual wire request.

Context

Inspired by any-llm#865 which addressed the same max_tokens vs max_completion_tokens issue in the Python port. The Go approach differs: rather than changing the base compatible provider, we keep max_completion_tokens as the correct default (per OpenAI spec) and let non-compatible providers opt into transforming their requests.

Test plan

• Wire-level tests using httptest.NewServer capture actual JSON bodies and assert field names
• OpenAI sends max_completion_tokens (not max_tokens)
• DeepSeek sends max_tokens (not max_completion_tokens)
• Mistral sends max_tokens (not max_completion_tokens)
• Mistral strips user and reasoning_effort from wire requests
• Full test suite passes with -race