Make image reading ops return tensors in CHW

[fmassa]

🚀 Feature

Motivation

All PyTorch operations (and our torchvision transforms) assume that images are in CHW format.
Traditionally, images are stored in HWC format, and for now we are returning contiguous tensors in HWC format.

I believe what we should do is to return non-contiguous tensors in CHW format, so that loading is as efficient as possible, and we can utilize channels_last format implementations for the operations (which are faster than the channels_first for a number of cases).

Pitch

Add .permute(2, 0, 1) after the image reading operations (in C++).