Add bitsandbytes: fine-tuned adapter was silently falling back to base model without it

Deploy: fix ZeroGPU RUNTIME_ERROR (@spaces.GPU decorator), sync prompt-format and greedy-decoding fixes from main training/eval work

Fix build: remove pinned gradio/transformers/peft/unsloth deps conflicting with HF Spaces platform gradio

Fix torch version for ZeroGPU compatibility

Deploy SQL Copilot: Gradio UI with Qwen model