fix(zerogpu): switch to gr.Blocks + add /run/synth_batch endpoint

bnb requirements

autoawq instead of bitsandbytes

add bitsandbytes for INT4

fix: pin huggingface_hub<0.26 (HfFolder), don't redeclare gradio (HF forces 4.44.0)

fix: pin gradio>=5.0 + huggingface_hub<0.30 (HfFolder removed in 0.30+)

initial: Qwen2.5-Coder-7B + Surrogate-1 v1 LoRA on ZeroGPU A10G