Shared memory optimizations for Gaussian rasterization

[matthewdcong]

1. Forward rasterization does not currently store features in shared memory. As the problem size becomes larger (more intersections with each Gaussian), the cost of an unconditional global load is outweighed by the shared memory reuse.
2. In addition, we cull loads for Gaussians with an opacity less than the threshold necessary for a Gaussian to be valid in the volume rendering pass. This optimization applies to the forward and backwards pass.

In profiling, this reduces a 17m 20s single-GPU reconstruction to 16m and 48s, leading to an approximately >3% speedup.

[harrism]

One concern.