feat(runtime): add Unsloth Gemma 4 support for E2B, E4B, and 26B-A4B

[slomin]

Closes #265

Summary

• Add first-class Potato support for Unsloth Gemma 4 GGUFs: E2B, E4B, and experimental 26B-A4B
• Model family detection, projector repo resolution, vision capability detection, and runtime env propagation
• Shell integration: mmproj handling, repo resolution, and candidate generation for all three variants
• Auto-switch to llama_cpp runtime when activating Gemma 4 models (ik_llama doesn't have gemma3n/gemma4 arch support yet — tracked in Feature Request: Gemma4 model support ikawrakow/ik_llama.cpp#1572)
• Strip Gemma 4 <|channel>thought tokens in the chat UI (upstream llama.cpp PEG parser bug, fix pending in Gemma 4 template parser fixes ggml-org/llama.cpp#21326)
• Fix cp same-file error in build_llama_runtime.sh universal profile packaging
• Refactored default_projector_candidates_for_model into a shared helper for both Qwen3.5 and Gemma 4
• Built and deployed upstream llama.cpp (commit a1cfb64) to the llama_cpp runtime slot on Pi

Test evidence

pytest: 649 passed in 10.45s
playwright: 144 passed (35.1s)

Pi QA

• ssd.local (Pi 5 8GB): E2B Q4_K_M — loads, vision works, projector auto-downloads ✓
• ssd.local (Pi 5 8GB): E4B Q4_0 — loads, inference works ✓
• potato.local (Pi 5 16GB): 26B-A4B IQ4_NL — loads (15.5GB RSS), inference at 2.76 tok/sec, vision with mmproj works ✓
• potato.local: Channel token stripping verified clean via Chrome MCP ✓
• Qwen3.5 regression: switched back, no issues ✓