Add Gemma 4 E4B thinking modes

[MilitantTurtle]

What changed

• Replaced the single automatic Gemma 4 E4B preset with explicit Thinking and Non-Thinking profiles.
• Kept Google's documented sampler: temperature 1.0, top-p 0.95, and top-k 64.
• Set Thinking to unlimited reasoning and Non-Thinking to a zero reasoning budget.
• Kept context and vision/projector selection out of the public preset because those remain machine-specific add-model choices.
• Added regression coverage using the official Google QAT GGUF filename.

Why

The previous reasoning: auto profile left mode selection dependent on runtime detection. Explicit modes make the intended quality/direct behavior predictable while retaining the same creator-sourced sampling defaults.

Validation

• python -m unittest discover -s tests -v — 28 passed
• python -m py_compile app.py setup.py
• node --check static/app.js
• node --check static/settings.js
• git diff --check
• Tracked-file privacy scan