fix(openai-adapter): map max_tokens to max_completion_tokens for DeepSeek reasoner

[RomneyDa]

Based on the work by @BurakBebek1 in #10478 — extracting just the max_completion_tokens fix as requested in review comments.

Summary

• DeepSeek Reasoner models require max_completion_tokens instead of max_tokens (similar to OpenAI's o-series models)
• Automatically maps the parameter when the API base includes api.deepseek.com or the model name includes deepseek-reasoner

Test plan

• Verify DeepSeek Reasoner models no longer fail due to max_tokens being sent
• Verify no regression for standard OpenAI or other provider logic
• Verify o-series handling still works as expected

---

Summary by cubic

Map max_tokens to max_completion_tokens for DeepSeek Reasoner models in the OpenAI adapter to prevent request failures. Applies when the API base includes api.deepseek.com or the model name includes deepseek-reasoner; standard OpenAI and o-series handling remain unchanged.

^{Written for commit 4f13740. Summary will update on new commits.}

[continue]

📝 Documentation Review: No docs update needed.

This PR implements an internal fix that automatically maps max_tokens to max_completion_tokens for DeepSeek Reasoner models. This is handled transparently in the OpenAI adapter—users don't need to configure anything differently or be aware of this parameter mapping. The fix simply makes DeepSeek Reasoner models work correctly out of the box, which is the expected behavior and doesn't require documentation.

[cubic-dev-ai]

No issues found across 1 file