Feature: Support think parameter for Ollama thinking models (Qwen 3/3.5, etc.)

[kishorgandham]

Feature Request

Problem

When using thinking models (e.g., qwen3:0.6b, qwen3.5:0.8b) via AxAIOllama, there's no way
to disable the thinking/reasoning phase. These models generate extensive <think>...</think> tokens
before producing the actual response, which significantly increases latency — often making requests
5-10x slower than necessary for simple tasks like classification.

What I've tried

1. thinkingTokenBudget: 'none' in forward() options — fails with:
   AxGenerateError: Model qwen3.5:0.8b does not support thinkingTokenBudget
   because ax-llm's model registry doesn't recognize custom Ollama model names as thinking-capable.

2. Ollama Modelfile with PARAMETER think false — Ollama doesn't support think as a Modelfile parameter.

3. Ollama's API does support "think": false in the chat request body, and the CLI supports
   /set nothink — but ax-llm doesn't pass this parameter through to Ollama.

Proposed Solution

Allow passing think: false (or the existing thinkingTokenBudget: 'none') through to Ollama's
API for custom models. Two possible approaches:

1. Pass thinkingTokenBudget: 'none' for all Ollama models — since Ollama model names are
   custom/user-defined, ax-llm can't know which ones support thinking. Instead of checking a
   registry, just pass "think": false in the request body when thinkingTokenBudget is 'none'.

2. Add Ollama-specific options — allow a pass-through for Ollama-specific API parameters
   (e.g., options: { think: false }).

Environment

• ax-llm: v19.0.13
• Ollama: v0.17.7
• Models tested: qwen3.5:0.8b, qwen3:0.6b

[dosco]

We can extend the OpenAI adapter to add the ability to strip out ... and add them to the right place in the openai response api