Add first-class retry configuration to `ChatOllama`

`ChatOllama` currently has no constructor-level retry-count option analogous to `max_retries` on many other LangChain chat integrations. Today: ```python from langchain_ollama import ChatOllama llm = ChatOllama(model="llama3", max_retries=3) ``` is accepted by Pydantic, but `max_retries` is not retained as a model field: ```python assert "max_retries" not in ChatOllama.model_fields assert getattr(llm, "max_retries", None) is None ``` Invoke-time kwargs can flow into the Ollama request payload, but that does not configure SDK/client request retries. This would be useful for applications that expose provider-agnostic retry configuration. For example, `deepagents-code` supports a `[retries]` config for retry-capable providers, but cannot safely register `ollama` because there is no first-class retry parameter today. Expected shape could be: - add `max_retries: int | None = None` to `ChatOllama` - apply the retry policy around both sync and async Ollama client calls - match the semantics used by other LangChain chat integrations that expose `max_retries` Related source context: - `ChatOllama` defines `client_kwargs`, `sync_client_kwargs`, and `async_client_kwargs` for HTTP client construction, but no retry-count field. - `_chat_params` forwards leftover invocation kwargs into the request payload, which is not the same as SDK/client retries.