Implement splats again
Now we have normals and splat radius information so can do a better job. It requires a single visibility step between dibr merge and attribute contributions kernel.
For each pixel, use a warp to search a window for possible points. For each found point do a splat intersection and use distance from point as a contribution value. Across all warps, accumulate plausible depth values before finally making a choice on an averaged nearest one. Perhaps then sync and do the attribute accum or leave that in another kernel and just output the depth for the pixel.