fix(2dgs): use ray-plane intersection depth instead of Gaussian center depth

[siddhu-pikachu]

Summary

This PR fixes 2DGS depth rendering so that the depth channel uses the ray-splat intersection depth instead of the Gaussian center depth.

The original issue was that rasterize_to_pixels_2dgs blended the packed Gaussian center depth in the last channel, which produced incorrect depth for tilted splats. In the repro from #863, this showed up as nearly constant depth across each splat instead of the expected gradient.

What changed

• Added a has_depth_channel flag through the 2DGS stack so the correction is only applied when the last color channel actually stores depth:
  • Python wrapper
  • torch bindings / schemas
  • C++ declarations / definitions
  • CUDA forward / backward paths
• Updated the 2DGS forward kernel to replace the packed center depth with the ray-splat intersection depth.
• Updated the 2DGS backward kernel so gradients for the depth channel flow through the intersection-depth computation instead of back into the packed center-depth color channel.
• Updated distortion / median-depth handling to use the same corrected depth quantity.
• Fixed the 2DGS binding/declaration mismatches needed to expose the new argument correctly.

Validation

I validated this with the issue #863 repro scene (two tilted splats) and compared the gsplat 2DGS output against the local Inria diff-surfel-rasterization reference backend.

Before this change:

• depth_2dgs was using Gaussian center depth and did not match the expected ray-plane/ray-splat intersection behavior.

After this change:

• depth_2dgs shows the expected spatially varying depth gradient.
• The output now matches the Inria reference visually on the repro scene.

Notes

There are still tiny residual rasterization-footprint differences between gsplat and the Inria backend, but those appear to be pre-existing and separate from the depth bug fixed here. This PR is focused on correcting the depth quantity used by the 2DGS path.

Closes #863