implement stacked spatial transforms as coordinates ops + resampling

[wyli]

stacked spatial transforms such as #107 #108 #109 #110 #111 could be implemented as computing multiple operations on a grid and doing one step resampling

This ticket looks for a "vanilla" non-gpu implementation

https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.map_coordinates.html
https://pytorch.org/docs/stable/nn.functional.html#grid-sample

[tvercaut]

I had a quick play with grid-sample and it performs quite well in comparison to other easily accessible alternatives for resampling:
https://colab.research.google.com/drive/17pI6xPGLTGcpGBulBrWZeNnBfLAwzqZT

Why not just stick to this for the resampling part (at least unless we find a faster or more feature rich implementation)?

scipy
10 loops, best of 3: 25.4 ms per loop

jax on device
10 loops, best of 3: 72.8 ms per loop

torch with Tesla T4
10000 loops, best of 3: 149 µs per loop

opencv cpu
1000 loops, best of 3: 1.3 ms per loop

cupy on device
100 loops, best of 3: 4.48 ms per loop

For the record, there is an interesting related discussion thread here:
pytorch/pytorch#24870

[wyli]

Thanks! looks very promising. we'll do some benchmarking for complete preprocessing pipelines, also need to consider the multi-gpu cases. @madil90