Input masks to suppress background/uncertain regions during splat optimization

[patriciogonzalezvivo]

This PR adds support for loading optional masks to the input pipeline in order to ignore unwanted regions during training. When --masks <FILES> are provided, the training routine uses them to weight or exclude those pixels from the splatting process, leading to cleaner reconstructions and less bleed-through from background or occluded areas.

Disclosure: this PR was made with the assistance of Claude Sonnet 4.6

What changed:

• Added mask-aware input handling for image data
• Propagated masks through the training/optimization path
• Adjusted the loss/weighting logic so masked regions contribute less or nothing to the splat fit
• Refined the masking behavior to reduce edge leakage around masked boundaries

How to use it:

opensplat \
  --images image_001.png image_002.png image_003.png \
  --masks mask_001.png mask_002.png mask_003.png

Test

Inputs:

Outputs:

https://superspl.at/scene/5b21b509

[pierotofy]

Approach is no good unfortunately. See #225 where current efforts are underway for a proper masking method.