Switch to FLUX.2-klein-4B: ungated, bf16 (no bnb, no ZeroGPU patch), 4-step default

Work around bnb-4bit shape check on ZeroGPU's fake CUDA device

show_error=True to surface runtime errors to the client

Split into stages (text encoder / denoise / preview decoder); ZeroGPU-safe loading; add torchvision

Load everything from ungated diffusers/FLUX.2-dev-bnb-4bit (4-bit transformer + text encoder, no token needed)

flux2 latent preview via Modular Diffusers streaming API (pipe.stream, diffusers PR #14159)

initial commit