GSoC: Incorporating point-based convolutions in TensorFlow Graphics (1/12)

[schellmi42]

As requested by @G4G , I split down the pull request #453 into smaller patches.

This is the first one.

1. point cloud class
2. regular grid computation
3. neighbor computation (ball query)
4. point density estimation
5. spatial sampling (poisson disk and cell-average)
6. point hierarchy class
7. 1x1 convolution and pooling layers
8. MCConv layer (incl. Monte-Carlo integration approximation method)
9. PointConv layer
10. KPConv layer (incl. constant summation approximation method and kernel point initialization)
11. example Colab-Notebooks
12. custom-ops package (CUDA-kernels)

Note:

The code is slightly different from the first PR #453. It's now possible to execute all functions in graph mode.