Leverage AMD FidelityFX state of the art radix sort library

[LifeIsStrange]

In the todo list there is:

Implement radix sort/cumsum by Taichi instead of torch, torch-taichi tensor cast seems only available on CUDA device. If we want to switch to other device, we need to get rid of torch.

I was looking at an alternative renderer repo (UnityGaussianSplatting) and they mention:

Splat sorting is done with a AMD FidelityFX derived radix sort.

AMD open source libraries under the gpuopen/fidelityfx umbrella generally are state of the art algorithms and use state of the art hardware optimizations/intrinsics. Moreover they generally performs well on Nvidia hardware.
https://gpuopen.com/fidelityfx-parallel-sort/
So it might be a good idea to use that