Tests for PyTorch api.
Before channel_last:
After channel_last: no pre-compile kernel, step time reduce 30%(compared to no-continus memory_format) Pytorch said they reduce 22% steptime compared to continus memory_format.
Usage: https://pytorch.org/docs/stable/tensor_attributes.html#torch.memory_format
Tutorial: https://pytorch.org/tutorials/intermediate/memory_format_tutorial.html
Pytorch benchmark: https://github.com/NVIDIA/apex/blob/ac8214ee6ba77c0037c693828e39d83654d25720/examples/imagenet/main_amp.py#L147C5-L147C57