Add support for channel-last `forward` in engines

**Is your feature request related to a problem? Please describe.**
As PyTorch already supports `channel last` data shape, channel-last support requires changing the layout of the model and data.
Need to add the support in our engines.

