speed up image encoder segment-anything-fast

[rbavery]

tracing a model can speed it up substantially and enable further optimization with tensorrt

examples of making the sam model trace-able: https://github.com/developmentseed/segment-anything-services/pull/25 https://github.com/facebookresearch/segment-anything/pull/326

and an example converting a Yolov5 model to tensorrt format and serving it with torchserve: https://github.com/rbavery/animal_detector/tree/master/mdv5app

[rbavery]

instead we can use https://github.com/pytorch-labs/segment-anything-fast/tree/main/segment_anything_fast

todo

• test torch.export and see if we can use portable .pt2 and if it supports dynamic shapes for image input
• if not, switch inference to use segtment-anything-fast library