Add uv-based installation workflow for faster setup

[MasahiroOgawa]

Summary

This PR adds a uv-based installation workflow as a faster alternative to conda, providing:

• install_env_uv.sh: Complete installation script supporting both CUDA 11.8 and 12.x
• activate_env.sh: Environment activation with automatic GCC/CUDA version detection
• pyproject.toml: Modern Python packaging with all dependencies declared

Key Features

• Faster installation: uv provides significantly faster dependency resolution and package installation compared to conda
• CUDA flexibility: Supports both CUDA 11.8 (older GPUs) and CUDA 12.x (modern GPUs including Blackwell/RTX 50 series)
• Automatic compiler detection: Finds compatible GCC version if system default is too new
• Simple activation: Single source activate_env.sh command sets up all environment variables

Installation Test

Successfully trained 3DGUT on NeRF Synthetic Lego dataset (500 iterations):

Metric	Value
PSNR	21.57 dB
SSIM	0.813
LPIPS	0.236

Test render output:

Usage

# Install (CUDA 12.8 default)
./install_env_uv.sh

# Or for CUDA 11.8
CUDA_VERSION=11.8 ./install_env_uv.sh

# Activate
source activate_env.sh

Test plan

• Fresh installation on Ubuntu with CUDA 12.8
• Training test on NeRF Synthetic dataset
• Verification of all package imports (torch, kaolin, tinycudann)

🤖 Generated with Claude Code