Model card: cut what the code repo already documents

Model card: cover the artifacts only, defer documentation to the code repo

Add prebuilt fp16 TensorRT backbone engine (AGX Orin, JetPack 6.2, TRT 10.3, 640x640)

Update latency figures; document fp16/CUDA-graph/TensorRT deployment

Fix --output usage: it is a directory, not a file path

Link the reproduction guide now that the code repository is public

Add AffKernel IIT-AFF checkpoint (R50vd stride-2 + deep supervision, seed 42) and model card

initial commit