Change depth from float64 to float32

[nepfaff]

It seems weird to me that depth used float64 while everything else used float32. Moreover, we specify that "Filepath points to a 16-bit or 32-bit depth image", so I don't seem a reason behind converting it to float64. Moreover, the depth prediction is also float32 (ds_nerf_depth_loss):

loss = -torch.log(weights + EPS) * torch.exp(-((steps - termination_depth[:, None]) ** 2) / (2 * sigma)) * lengths

Here steps is float32 while termination_depth is float64. This PR changes it to also be float32.

Making this change leads to a small speedup.

Before change (float64):

After change (float32):

This is towards #3446.

[brentyi]

Yeah, this looks unintentional. Thanks!!