torchao int8 weight-only + flash-linear-attention (bnb int8 was ~4 tok/s)

int8 (bitsandbytes) on large slice; scale GPU duration to max_new_tokens

Add app.py