Add tool_choice parameter to chat() (mirror server-side ollama#11171)

Observation

The chat() function and underlying ChatRequest model in ollama-python do not expose a tool_choice parameter, even though several upstream endpoints already handle it:

• OpenAI-compat endpoint in ollama/ollama server (openai/responses.go): ToolChoice is parsed off the request, with TODO comments noting partial support.
• Anthropic-compat endpoint (anthropic/trace.go + anthropic/anthropic.go): ToolChoice is wired through.
• Native /api/chat: tool_choice is not a request field today; tracked upstream at ollama/ollama#11171 ("Add support for tool_choice set to any to require a model to use a tool").

Empirical behavior today

import ollama

# Test 1: tool_choice as kwarg → hard rejection
ollama.chat(
    model="llama3.1:8b",
    messages=[{"role": "user", "content": "Hello"}],
    tool_choice="auto",
)
# TypeError: Client.chat() got an unexpected keyword argument 'tool_choice'

# Test 2: tool_choice via options dict → silently accepted but not wired
# (options is passed through as model parameters; tool_choice has no effect there)
ollama.chat(
    model="llama3.1:8b",
    messages=[{"role": "user", "content": "Hello"}],
    options={"tool_choice": "auto"},
)
# OK, no error, but tool_choice is not interpreted by the server

ChatRequest.model_json_schema() confirms there is no tool_choice field. Current fields: model, stream, options, format, keep_alive, messages, tools, think, logprobs, top_logprobs.

Why it matters

Practitioners porting code from OpenAI's or Anthropic's SDK to ollama-python expect a tool_choice parameter (to force tool use, restrict to a specific tool, or disable tools). Today the only available control is omitting tools entirely, which is coarse.

Concrete suggestion

When the native /api/chat server endpoint adds tool_choice (per ollama/ollama#11171), expose the same parameter on ollama.chat() and ChatRequest:

# Standard cross-vendor values
tool_choice: Literal["auto", "required", "none"] | dict | None = None

Until then, this issue serves as a tracking item for the Python client mirror so it lands in the same release window as the server-side feature.

Workarounds

• For "must use a tool" semantics, append a directive to the system prompt and use temperature=0.0 (verified to produce reliable tool calls on Llama 3.1 8B).
• For "must not use a tool", omit the tools argument entirely.

Happy to send a PR once the server-side direction is settled.

Start with the chat() entry point and the ChatRequest model, then inspect ChatRequest.model_json_schema() to confirm the current request surface. Track ollama/ollama#11171 and the mentioned OpenAI-compat and Anthropic-compat paths before proceeding; done means the Python client mirrors the settled native server behavior for tool_choice.