Fail fast on hard Gemini quota instead of spamming every key

Switch try-on and garment analysis to external image APIs

feat: default STORAGE_MODE to cloudinary and force real CatVTON execution on CPU

feat: add CPU simulation fallback mode for tryon to prevent hangs/OOM on CPU spaces

perf: optimize CatVTON on CPU with float32 and 20 steps, increase frontend polling to 10 min

chore: move CatVTON inside backend and resolve paths dynamically without binary resources

feat: integrate local CatVTON virtual try-on and secure env files

Initialize production-grade FastAPI backend with models, routes, celery workers, and test suite