Make PixelSampler the depth-nerfacto default

[nepfaff]

The default pixel sampler for depth-nerfacto is PairPixelSampler. However, this is very slow due to mask erosion (so slow that it is basically unusable). I managed to speed it up a little here #3452. However, it is still much more efficient to use PixelSampler. Moreover, PairPixelSampler is only needed when using SPARSENERF_RANKING as the depth loss, but the default value is DS_NERF, which doesn't require PairPixelSampler. Hence, with the default depth loss, most people will be much better off with PixelSampler instead of PairPixelSampler.

I tested with both depth and masks using 50 images of resolution 1920x1440. Note, that PairPixelSampler is only slow when training with masks due to the need for mask erosion.

With PairPixelSampler:

With PixelSampler:

Closes #3446

[akristoffersen]

This makes sense to me. Approving.

It's a bit unfortunate that there's no obvious sign to the user that PairPixelSampler should be used when using SPARSENERF_RANKING. I believe it would silently fail if that depth loss was used without the pair pixel sampler. Could you add to the comment under the loss type in DepthNerfactoModelConfig to note this?

[nepfaff]

Added a comment. It would be best to throw an exception in DepthNerfactoModel but that doesn't seem to be possible with the current structure.