`RandAffine`, `RandFlip`, `RandRotate` with GPU-based `torch.tensor`

[wyli]

Is your feature request related to a problem? Please describe.
this is as part of #2231 would potentially accelerate the current test-time augmentation pipeline, when combined with https://github.com/Project-MONAI/tutorials/blob/master/acceleration/fast_training_tutorial.ipynb

see also: #2913

cc @diazandr3s

[diazandr3s]

This is great! It will help a lot in accelerating TTA used in the MONAI Label Apps

[wyli]

addressed by #2922 #2945