Remove PEFT adapter files, switch app.py to load merged model

Switch to Gradio + ZeroGPU (Docker SDK not available on this account)

Add the real BLOOMZ generator inference API