[codex] Normalize Azure OpenAI max token aliases

[54sun]

Summary

• normalize max_completion_tokens to max_tokens before Azure OpenAI provider execution
• add regression coverage for Azure OpenAI factory calls

Root cause

Newer model helpers emit max_completion_tokens for gpt-5.* models, but DeepTutor's Azure Responses provider accepts max_tokens and maps that internally to max_output_tokens. The alias leaked through and caused Azure OpenAI calls to fail before reaching the API.

Validation

• python -m py_compile deeptutor/services/llm/factory.py tests/services/llm/test_factory_provider_exec.py
• live Docker backend LLM and embedding connection tests passed