Fusing PTv3 point serialization into custom CUDA kernel

[ChristianSchott]

Hello,
thanks for sharing this great work!

I wanted to link this here since LitePT focuses on efficiency.

Original issue: Pointcept/Pointcept#578
CUDA implementation: https://github.com/ChristianSchott/point_serialization_cuda

Using the base PTv3 model, switching the serialization step to CUDA gives about a ~22% speedup. On LitePT, the gains are slightly smaller since not all downsampling steps involve serialization, but LitePT-S still sees around 15% improvement.

[ywyue]

Hi @ChristianSchott, thanks for pointing it out! Would you like to submit a PR so that users can switch between your CUDA implementation and the original one?