Update docs to recommend torch 2.1

[brentyi]

I couldn't get gsplat to build automatically from a PyPI install on any of my servers with torch 2.0.1. gsplat worked fine when installing from the git repo.

The root cause was that my CUDA libraries, when installed via conda, needed to be linked from $CUDA_HOME/lib and not $CUDA_HOME/lib64. This is fixed in 2.1: pytorch/pytorch#101285

[brentyi]

For posterity, the compile error I was running into was:

[5/5] c++ bindings.cuda.o forward.cuda.o backward.cuda.o ext.o -shared -L/home/brent/miniconda/envs/nerfstudio/lib/python3.8/site-packages/torch/lib -lc10 -lc10_cuda -
ltorch_cpu -ltorch_cuda -ltorch -ltorch_python -L/home/brent/miniconda/envs/nerfstudio/lib64 -lcudart -o gsplat_cuda.so 
FAILED: gsplat_cuda.so                                                                                                                                                 
c++ bindings.cuda.o forward.cuda.o backward.cuda.o ext.o -shared -L/home/brent/miniconda/envs/nerfstudio/lib/python3.8/site-packages/torch/lib -lc10 -lc10_cuda -ltorch
_cpu -ltorch_cuda -ltorch -ltorch_python -L/home/brent/miniconda/envs/nerfstudio/lib64 -lcudart -o gsplat_cuda.so                                                      
/usr/bin/ld: cannot find -lcudart: No such file or directory                                                                                                           
collect2: error: ld returned 1 exit status                                                                                                                             
ninja: build stopped: subcommand failed.                                                                                                                               

This only happened from the JIT compile that happens when running ns-train gaussian-splatting, and not the install-time setup.py / CppExtension compile. (which follows a different code path in torch)

[tancik]

Seems reasonable enough

[Guangyun-Xu]

Does the Dockerfile also need to be updated?

[ichsan2895]

Strangely, FYI,
My Runpod environment run Gsplat successfully with Torch==2.0.1. I will stay in this version despite dump version in project.toml.