[3D Gaussians] Performance improvements

[Meakk]

While the performance is acceptable on high-end GPUs, it's still a bit slow when there are many splats.

Currently, 3D Gaussians rely on Bitonic Sort to sort points from back to front for proper alpha compositing. However, the state of the art is the "Onesweep" radix sort, so it should be faster.

Also, it would be more performant to sort from front to back and stop rendering fragments when we reach an opacity threshold.
Changing the blend function to support front-to-back is easy, but how can we tell OpenGL to early exit when the current fragment is opaque?