Recognise (if only to strip/ignore) <think> tags

[lee-b]

With Qwen3 (and other models, I'm sure), tags appear in the output. These should be hidden, or totally stripped from the output, as they're basically an internal tool to help the LLM create a useful final response, when it gets to the final response.

One option might be to log the thinking and the response in debug mode, but otherwise just drop the think tags.

[p-e-w]

Yes, that makes sense.