fix(llm): stop putting format/stream inside Ollama options - #5561

Summary

• format and stream passed through ChatOllama(ollama_options=...) are not model options. They are top-level chat() parameters, and forwarding them is what produced the truncated structured JSON in Bug: Invalid Json error in browser-use while using ollama vision models, I have used upto 27B paramete models. #5017.
• Unwrap markdown json fences before model_validate_json, which Ollama vision models often add around the payload.

Fixes #5017.

Test plan

• uv run pytest tests/ci/models/test_llm_ollama.py -q failed on main, then passed after the change
• CI tests / lint on this PR

Summary by cubic

Fixes #5017 by stopping format and stream from being passed inside Ollama ollama_options; they are top-level chat() parameters and caused structured output to break.

• Moves think, logprobs, top_logprobs, and keep_alive to top-level chat() parameters as well.
• Strips markdown JSON fences that Ollama vision models often wrap around responses before parsing.
• Surfaces invalid JSON as a ModelProviderError instead of a raw validation error.
• Adds unit tests covering option splitting, fence handling, and truncated JSON.

^{Written for commit ed5f467. Summary will update on new commits.}