Allow training and rasterizing on a full image background in addition to flat colors

[brianshaler]

I imagine random_background with a flat color works for most use-cases with masked inputs. I had an instance where I needed to train on opaque RGBs that were pre-composited with a known background that I could also pass in as the background when training.

This is probably more generally useful for rasterization.

Try it out with noise:

background = torch.rand((3, viewpoint_camera.image_height, viewpoint_camera.image_width), dtype=torch.float32, device="cuda")

[aixiaodewugege]

Thanks for your brilliant work. Does the original version not support a custom background? Is there any document explain it?