Add experimental GEMM forward rasterization backends for 3DGS and 2DGS

[zhubowen1]

This PR adds experimental GEMM-based forward rasterization backends for both 3DGS and 2DGS.

Background

The existing rasterization pipeline relies on the original per-pixel forward rasterization kernels. This PR introduces an alternative forward backend designed to accelerate rasterization by reformulating key parts of the computation into GEMM-style, tensor-core-friendly workloads.

Changes

• Added a GEMM-based forward rasterization kernel for 3DGS in RasterizeToPixels3DGSFwd.cu.
• Added a GEMM-based forward rasterization kernel for 2DGS in RasterizeToPixels2DGSFwd.cu.
• Added shared low-level GEMM raster helpers in GemmRasterUtils.cuh.
• Added explicit backend selection plumbing through Python, C++, and CUDA: rasterize_fwd_impl="gemm"

Evaluation Setup

I evaluated the GEMM backend on seven real scenes: garden, bicycle, stump, bonsai, counter, kitchen, and room.
These scenes cover both outdoor and indoor settings and vary in scene complexity. I report rendering quality measured by PSNR, and performance measured by the latency of the forward rasterization kernel only.

Performance Summary

Across these seven scenes, the GEMM backend consistently reduces the latency of the forward rasterization kernel on both RTX 3090 and Jetson AGX Orin, while maintaining nearly identical PSNR.
Note that the reported latency refers only to the single forward rasterization kernel, rather than the full end-to-end rendering pipeline.

Platform	Backend	Avg legacy latency	Avg GEMM latency	Speedup	Latency reduction
RTX 3090	3DGS	2.61 ms	2.00 ms	1.33x	23.30%
RTX 3090	2DGS	4.80 ms	3.29 ms	1.47x	31.30%
Jetson AGX Orin	3DGS	51.79 ms	29.18 ms	1.87x	43.70%
Jetson AGX Orin	2DGS	87.89 ms	51.84 ms	1.70x	41.00%

Quality

The GEMM backend preserves rendering quality closely compared with the legacy backend.

Backend	Avg PSNR change	Max absolute PSNR change
3DGS	-0.019 dB	0.041 dB
2DGS	-0.004 dB	0.022 dB

Known Limitations

• Forward-only support
  The GEMM rasterization backend currently supports forward inference only. Backward/gradient computation is not implemented yet.
• Half-precision computation
  The implementation relies on Tensor Core (MMA) operations in FP16. While this enables significant speedup, it may introduce minor numerical differences compared to the FP32 legacy backend.
• Fixed tile size
  The current implementation only supports a tile size of 16×16.
• Limited channel configuration
  Only CDIM = 3 is supported in the current version.
• Hardware requirement
  Requires GPUs with Tensor Core support (SM 7.5 and above).

[zhubowen1]

This GEMM-based rasterization backend is related to our recent work:

• Paper: GEMM-GS: Accelerating 3D Gaussian Splatting on Tensor Cores with GEMM-Compatible Blending
  https://arxiv.org/abs/2604.02120

• Reference implementation:
  https://github.com/shieldforever/GEMM-GS