fix: use torch.cuda.amp.autocast() for mixed-precision encoding

perf: use FP16 precision on CUDA and SQLite executemany

feat: auto-detect GPU/MPS in generate_embeddings.py

feat: model auto-detection, startup warmup, RRF k=40, CAMeL-BERT default, switch_model helper

Deploy Fiqh.ai FastAPI Backend with Database and Git LFS