Add structure channel for rendering with masks

[eigenvivek]

output of drr.forward should be B C H W where C is the number of structures in drr.subject.mask (C=1 if subject has no mask).

current idea is

• use the same indexes for the volume on the mask, ie channels = torch.take(mask, idxs) with shape B (HW) N
• onehot channels for shape B C (HW) N
• multiply channels and voxels, nansum over N, and make img with shape B C (HW)

current limitation: onehot is memory intensive for C = 118 (ie, totalsegmentator)

[eigenvivek]

onehot can be bypassed: https://stackoverflow.com/questions/78323859/broadcast-pytorch-array-across-channels-based-on-another-array