Is your feature request related to a problem? Please describe.
Hi @rijobro ,

Thanks for providing the TTA module since MONAI 0.5.
@dongyang0122 @holgerroth and I discussed the TTA module for evaluation cases, and found several things should be updated:

1. All the invert transform logic in TTA can be replaced by Invertd transform directly now.
   And Invertd transform provides more features.
2. The post transform example in the doc-string is not decollate logic.
3. Better to ensure determinism for the random transforms in the example of doc-string.
4. It copy between Numpy and Tensor several times because we didn't support Tensor in transforms and don't have the proper numpy <-> tensor conversion before:
   https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/test_time_augmentation.py#L188
   https://github.com/Project-MONAI/MONAI/blob/dev/monai/data/test_time_augmentation.py#L190
   It will also cause numpy error if the transforms inverse() outputs Tensor.
   Should be updated to support NdarrayOrTensor directly.
5. Some doc-strings exceed 120 chars.