Improve inference speed on sagemaker serverless while preserving accuracy

[rbavery]

Goal: get to 2-3 seconds per image.

according to yolov5 benchmarks and our past tests, inference speeds are possible, but with potential requirements to compromise on image size: ultralytics/yolov5#6613

PR: ultralytics/yolov5#6613

[nathanielrindlaub]

We got to ~3.5 seconds per image with the ONNX-compiled model deployed on Sagemaker Serverless - so chalking that up to a win for now (see #105). Nice job @rbavery thanks for your persistence on this!!