Testing by reloading the model on __call__ with quantization in case GPU is not available during __init__

Added back path parameter in case it has to be in the signature

Added handler.py and requirements.txt

Duplicate from llava-hf/llava-1.5-7b-hf