CUDA implementation of higher-order image resampling (e.g. cubic spline)

[tvercaut]

Is your feature request related to a problem? Please describe.
Pytorch's grid-sample function provides a GPU implementation of linear interpolation for image redsampling. Registration networks and warping-based augmentation layers may benefit from higher-order image resampling which is currently not available in either pytorch or MONAI.

Describe the solution you'd like
Implement a CUDA-based higher-order interpolation layer allowing for a backward pass (i.e. differentiable).

Describe alternatives you've considered
N/A

Additional context
NiftyNet ported the code from NiftyReg for that purpose: https://niftynet.readthedocs.io/en/dev/niftynet.contrib.niftyreg_image_resampling.html

This issue spins from #95 and relies on #785 for enabling a propoer packaging of such feature.

[tvercaut]

Copying some notes from #1452 to ease the discussion.

The nitorch-based code in MONAI (https://github.com/Project-MONAI/MONAI/tree/master/monai/csrc/resample) only partially adresses the need. It seems we have two options:

• increase the scope of the current implementation and also include pre-filtering (e.g. based on https://github.com/DannyRuijters/CubicInterpolationCUDA)
• switch to a more complete implementation (e.g. based on niftyreg as done for NiftyNet https://github.com/NifTK/NiftyNet/tree/dev/niftynet/contrib/niftyreg_image_resampling)

[vikashg]

closing as the issue was closed earlier