CUDA V2

[liruilong940607]

I dramatically rewrite the CUDA part to make the code runs more efficient, with cleaner code structure, and some extra functionalities.

New features:

• Support batch rendering in one pass: N cameras -> images with shape (N, H, W, C). This allows for sampling evenly distributed cameras during every training iteration, thus could potentially lead to better convergence. With this, we probably don't need gradient accumulation anymore. Also, for rendering GS, if some latency can be allowed, we can use batched rendering to further boost FPS.
• Supports fully PyTorch Auto-Grad: Provides a version of the rasterization & projection implementation that the gradients are fully managed by PyTorch Auto-Grad (see _rendering in experimental/cuda/__init__.py). This allows for easier try-out for the new ideas as the modifications can probably done purely on the python side. It could also serves as a gradient checker for the CUDA implementation.
• Fused implementation with verified Gradients: Every gradient in this PR is carefully checked against PyTorch Auto-Grad. On the other hand, the current GSplat implementation has flaws (errors) in the gradients including -- 1. FoV clip is applied in the forward pass but not the backward pass. As a result, the gradients are incorrect on some GSs. -- 2. The backward suffers from a numerical issue caused by using the remained transmittance output from the forward pass because the remained transmittance can be very small (~1e-4). Double precision instead of float should be used for it.
• Faster FWD and BWD (esp. in ND case): See table below for details.

Gradient Diff results again PyTorch Auto Grad (code in experimental/test.py):

	viewmats	scales	quats	means
GSplat	42.8834	2e-4	5e-4	4e-3
This PR	4e-3	2e-7	5e-6	3e-6

Runtime Performance (in ms) for Forward (code in experimental/profile.py):

Forward, Channel=1	Batch=1	Batch=8	Batch=64
GSplat	1.492	9.921	76.295
This PR	1.146	3.593	28.948

Forward, Channel=32	Batch=1	Batch=8	Batch=64
GSplat	5.078	32.824	264.146
This PR	1.930	7.773	64.461

Runtime Performance for Backward (disable viewmats gradients):

Backward, Channel=1	Batch=1	Batch=8	Batch=64
GSplat	1.594	10.600	84.395
This PR	1.562	7.409	60.402

Backward, Channel=32	Batch=1	Batch=8	Batch=64
GSplat	10.242	80.935	653.627
This PR	7.850	43.392	353.733

Runtime Performance for Backward (enable viewmats gradients):

Backward, Channel=3	Batch=1	Batch=8	Batch=64
GSplat	2.721	16.629	130.977
This PR	2.183	10.697	85.817