fix the error of illegal memory access caused by cub

[ShunChengWu]

Fix an issue caused by cub.
From my own test, the problem happened when you call cub related functions, such as InclusiveSum Here https://github.com/graphdeco-inria/diff-gaussian-rasterization/blob/main/cuda_rasterizer/rasterizer_impl.cu#L281.
The problem is that the copy process does not go all the way. From my test, the copy somehow stopped at around 80K. After I tried to bypass that function by using std::inclusive_scan. However, illegal memory access happened somewhere in the backward pass.
Afterward, I found that a similar issue kind of happened in custom-built Pytorch functions (NVIDIA/thrust#1341) and found the solution here pytorch/pytorch#52663 (comment).

After the test, I used the modified code to train Gaussian Splatting on the Blender Lego dataset, and it worked like a charm.

Related issues:

• CUDA error while running gaussian-splatting#23
• RuntimeError: CUDA error: an illegal memory access was encountered gaussian-splatting#41
  Potentially related issue:
• CUDA error in backward. gaussian-splatting#81

[N-Demir]

Strangely it didn't work for me either. Potentially it matters what cuda/torch version you're on? I've tried cuda11.8/torch1.13 and cuda12.4/torch2.4. And this is for using the 3dgs_accel branch

update: I had just messed up some code!

[HenryLyu]

Thank you very much! Your solution is awesome