Profiling and Improving the PyTorch Dataloader for high-latency Storage: A Technical Report