fix(backend): restore real trained MobileNetV2 and EfficientNet-B0 neural network prediction pipeline

fix(backend): bypass heavy torch loading in Vercel serverless environment for zero OOM crashes

fix(backend): add numpy tensor guards in disease and soil models for serverless environment

fix(backend): add resilient fallbacks for disease, soil, and image preprocessing in serverless environment

fix(vercel): optimize serverless dependencies and crop model fallback to pass 500MB function size limit

feat: replace legacy frontend, integrate agriculture-ai-frontend with backend, update CORS, and clean project structure

Add trained ML models (crop, disease, soil), tools, setup scripts, and updated .gitignore

first commit