Disable thinking in LLMs so prompt expansion will work properly with thinking models

[shanedk]

This is my first ever PR, so be gentle! 😊

Summary

A bug fix to disable thinking on text LLM models so prompt expansion will work properly with thinking models.

Related Issues / Discussions

Closes #9379

QA Instructions

1. Try a thinking model. I was using Goekdeniz-Guelmez/Josiefied-Qwen3-1.7B-abliterated-v1
2. Type in a prompt
3. Use it to expand a prompt
4. Notice that you just get thinking tokens, no prompt
5. Apply this change
6. Try it again
7. Notice that you get a nicely expanded prompt now

Merge Plan

This is just adding an extra parameter, so I'm unaware of any such issues.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[Pfannkuchensack]

PR #9380 — Disable thinking in LLMs

• No test coverage, and the existing tests cannot catch a revert. Both TextLLMPipeline suites drive the pipeline with MagicMock tokenizers that swallow **kwargs, and every assertion reads only call_args[0][0] (the messages list). Reverting this PR leaves the suite green. A hermetic test is cheap: invokeai/backend/qwen3/qwen3_tokenizer.py:27 exposes load_bundled_qwen3_tokenizer() and the real Qwen3 chat template ships in invokeai/backend/qwen3/tokenizer/tokenizer_config.json — no network, no weights, no GPU. Assert the rendered prompt ends with <think>\n\n</think>\n\n.
• The retry branch at invokeai/backend/text_llm_pipeline.py:63-65 has no coverage of any kind. Add a fake tokenizer that raises ValueError("System role not supported") on the first call and assert the second apply_chat_template call also carries enable_thinking=False.
• Pre-existing, but inside the block being edited: invokeai/backend/text_llm_pipeline.py:74 re-tokenizes the rendered template with add_special_tokens at its default True. transformers' own tokenize path passes False precisely because the template already emits its control tokens — this duplicates BOS for Gemma-family models. Qwen has no BOS, which is why Qwen-based manual testing could not surface it.
• No docs and no user-facing signal about which models the fix actually covers. All three shipped TextLLM starters are non-thinking models, so the affected population is entirely user-installed.

[shanedk]

It happened with almost every other model I tried: Qwen/Qwen3-0.6B and 1.7B (and variants), google/gemma-3-1b-it (and variants), and a few others.

I haven't had any issue with Qwen3 or gemma3-based models since I added this, and it doesn't seem to have created any problems with Llama-3.2 or SmolLM.