Add post-training splat compression

[92401]

Motivation

Spirula Studio makes it practical to train detailed Gaussian models, including
on relatively modest hardware. As the splat count grows, however, the resulting
models can become costly to store, transfer, and browse interactively on
devices with tighter VRAM and rendering budgets.

This PR focuses on the post-training workflow. It provides a reproducible way
to derive a smaller model from a completed checkpoint, measure the rendering
change introduced by compression, and evaluate the compressed model against
the source images before choosing a deployment trade-off.

Method

The compression workflow combines several simple and interpretable operations.

1. Contribution-based pruning

   The model is rendered from a representative set of training cameras. Each
   splat accumulates its alpha- and transmittance-weighted rendering
   contribution, and only the highest-scoring splats are retained.

2. Spherical-harmonics reduction

   Higher-order view-dependent colour coefficients are removed while preserving
   the geometry and lower-order appearance representation.

3. Opacity-threshold pruning

   Nearly transparent splats can be removed as a lightweight standalone
   pruning method or as an additional filter.

The commands operate directly on existing checkpoints and derived PLY files;
they do not require retraining.

CLI

This PR currently adds a command-line interface only; a GUI integration is
left for follow-up work.

splat prune

• Contribution-based pruning with --keep-fraction
• Opacity-threshold pruning with --min-opacity
• Spherical-harmonics degree reduction with --sh-degree

splat compare

• Renders the original and compressed models from identical dataset cameras
• Reports L1, PSNR, and SSIM to quantify compression-induced render drift

splat evaluate

• Renders a model at dataset camera poses and compares it against source images
• Reports raw and colour-corrected L1, PSNR, and SSIM

Model replay supports 3dgs, mip, and 3dgut. For standalone PLY files, the
--primitive option can override the primitive saved in config.json.
3DGUT is validated with a 3,000-step half-resolution Garden / pinhole smoke run. With the same 60% contribution pruning and degree-3 to degree-2 SH reduction, the model changed from 105.7 MB / 446,978 splats to 41.9 MB / 268,187 splats. On 20 source cameras, source-image PSNR dropped by 0.0774 dB and SSIM by 0.007129; original-versus-compressed renders measured 37.5592 dB PSNR and 0.987195 SSIM.

Benchmark

All compressed models retain 60% of contribution-ranked splats and reduce
degree-3 spherical harmonics to degree 2. Each result uses 20 source cameras.

Dataset / camera	Primitive	Original	Compressed	PSNR drop	SSIM drop
Garden / pinhole	Mip	248.0 MB	98.4 MB	0.5289 dB	0.013090
indoor6k / equirectangular	3DGS	682.2 MB	270.7 MB	0.5678 dB	0.006574
indoor6k / dual fisheye	3DGS	743.6 MB	295.0 MB	0.0803 dB	0.001781
Garden / pinhole	3DGUT (3k steps, 1/2 resolution)	105.7 MB	41.9 MB	0.0774 dB	0.007129

The indoor6k data was captured with an Insta360 X6. The dual-fisheye result
uses frames extracted from the original, unstitched INSV files, while the
equirectangular result uses frames extracted from an MP4 exported by Insta360
Studio. These private datasets are used only for evaluation and are not
included in this PR.

Validation

Tested on the latest upstream master commit:

• CUDA build completed successfully.
• Vulkan build completed successfully.
• splat_prune_test passed on both CUDA and Vulkan backends.
• Generated kernel instantiations were checked with
  python tools/codegen/generate_kernel_instantiation.py --check.
• Private-path and i18n checks passed.
• git diff --check upstream/master...HEAD reported no whitespace errors.