Fixed ABI version caused issues when using CUDA 12.8

[MrZoyo]

Hello.

This pull request introduces several improvements and fixes across mesh processing, rasterization, and C++ extension compatibility.

I need to clarify: some changes were made for my personal testing convenience. For instance, since my GPU VRAM configuration is low, I had to rewrite nvdiff_rasterization to use chunked processing to reduce VRAM overflow. The clean_convert_mesh.py script was created for my personal convenience in obtaining obj/glb formats.

These changes have only been tested on my specific device and configuration, without extensive verification.

Modifications to header files to enforce ABI=1 may adversely affect users of older versions. I apologize for not having older hardware available for testing.

These concerns are also why I didn't submit a PR earlier. I request that your team test my modifications whenever feasible and feel free to make any changes.

===========================================
Mesh Processing Utility:

• Added a new standalone script clean_convert_mesh.py for cleaning, simplifying, and converting mesh files (PLY/OBJ/GLB) with options for component filtering and triangle count reduction. This script leverages pymeshlab and optionally trimesh for robust mesh handling and exports.

Rasterization Improvements:

• Introduced a new nvdiff_rasterization function in mesh.py that supports chunked rasterization for large meshes (controlled via the MILO_RAST_TRI_CHUNK environment variable) and fixes a critical compatibility issue by ensuring the ranges argument for nvdiffrast is always on the CPU.
• Updated the mesh class initialization to allow selection of nvdiffrast backend (gl or cuda) via the NVDIFRAST_BACKEND environment variable, improving flexibility for different hardware setups.

C++ Extension Compatibility:

• Added a new header force_abi.h to force use of the C++11 ABI for GCC's libstdc++ in the tetra_triangulation extension, preventing runtime linker errors with recent PyTorch versions. This header is now included as the first line in all relevant C++ sources.
  CUDA and Header Hygiene:

• Added missing standard headers (<cstdint>, <cstddef>) to several rasterizer implementations for improved portability and reliability.

• Fixed CUDA pointer usage in spatial.cu by switching from .data<float>() to .data_ptr<float>() for compatibility with newer PyTorch versions.

• Added <cfloat> to simple_knn.cu and removed redundant macro definition for better CUDA compilation hygiene.