fix: #2823 AnyLLM reasoning extraction for iterable vLLM/any-llm Reasoning objects

[ziqi-cn]

Summary

Fix AnyLLMModel reasoning extraction on the chat-completions path for any-llm Reasoning objects.

This was observed with a vLLM OpenAI-compatible endpoint, where reasoning deltas reached AnyLLMModel as values like Reasoning(content="用户"). The previous implementation flattened generic iterables before checking .content, .text, and .thinking, so iterable Pydantic objects could be normalized incorrectly as content用户 instead of 用户.

This change makes the helper prefer reasoning attributes first, then fall back to iterable flattening.

Test plan

• Added a regression test covering Reasoning(content="用户")
• Ran uv run pytest tests/models/test_any_llm_model.py -q
• Ran make format
• Ran make lint

Issue number

Closes #2823

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation
• I've run make lint and make format
• I've made sure tests pass