Release v0.4.8

Features

• Implement standard and differentiable Marching Tetrahedra Grid operator (marching_tetrahedra_grid) for rapid voxel surface extraction with 6x tetrahedra subdivision per cube.
• Implement CUDA kernel execution and autograd backward pass for grid-based tetrahedra marching.
• Add PyBind11 C++ API bindings and lookup tables for tetrahedra subdivision and triangulation.
• Export data, io, and marching_tetrahedra_grid operators in top-level package bindings.

Refactors & Build

• Refactor triangle quality metric in mode 2 to compute minimum over maximum edge length ratio (l_min / l_max).
• Monkey-patch PyTorch CUDA version verification in setup.py to support building extensions with mismatched local nvcc compilers.
• Bump project release version to 0.4.8.
• Clean up deprecated experimental benchmarks and obsolete evaluation artifacts from marching cubes and marching tetrahedra experiments.

Examples

• Add end-to-end evaluation pipeline demonstrating ultra-fast grid-based Marching Tetrahedra surface extraction on real 3D asset data (StanfordBunny), achieving >8x speedup over reference Marching Cubes on GPU.