Fix local model base handling

Update download gate by backend

Simplify model download labels

Gate first model download behind start button

Add first-load download progress indicator

Fix static tokenizer loading from Hugging Face model repo

Deploy static FLUX.2 WebGPU app