DX: Reliable imports (PYTHONPATH) + robust CUDA build hints in setup.sh

[Mars122]

Title: DX: Make runtime imports reliable and harden setup for CUDA extension builds

Summary

This PR improves the out-of-the-box runtime experience by ensuring local module imports work and by making the CUDA extension build path more robust:

• Scripts now export PYTHONPATH to include the repo root and src/ so import src.* works regardless of environment.
• setup.sh sets CMAKE_PREFIX_PATH from Torch, CUDA header paths (CPATH), and CUDA_HOME, and builds RaDe-GS submodules with --no-build-isolation when available. ffmpeg install is guarded.

Changes

• scripts/infer.sh: export PYTHONPATH with $(pwd):$(pwd)/src
• scripts/train.sh: export PYTHONPATH with $(pwd):$(pwd)/src
• settings/setup.sh:
  • Export CUDA_HOME and CPATH
  • Set CMAKE_PREFIX_PATH via torch.utils.cmake_prefix_path when available
  • Clone RaDe-GS with --recursive --depth 1
  • Install diff-gaussian-rasterization (and simple-knn if present) with --no-build-isolation
  • Guard ffmpeg apt-get under sudo/apt presence

Motivation

Some users encounter ModuleNotFoundError: No module named 'src' depending on how scripts are launched, and CUDA extension builds may fail without explicit Torch/CUDA path hints. This PR makes both paths more reliable without changing core logic.

Notes

• No changes to the core training/inference code.
• All changes are optional and best-effort; they avoid failing on environments where certain tools are unavailable.

License

• Shell script adjustments only.