docs: add cuda-version note to conda install instructions

[evasnow1992]

The default conda install nvmolkit may select cuda-version=13, causing PyTorch to fall back to CPU-only on systems with CUDA 12.x drivers. Add a note instructing users to specify cuda-version to match their system.

For addressing #101

[greptile-apps]

Greptile Summary

This PR adds a helpful note to the conda installation instructions in README.md, explaining that the default conda install nvmolkit may resolve cuda-version=13 on systems with CUDA 12.x drivers, causing PyTorch to run in CPU-only mode. It introduces a workaround command pinning cuda-version=12.6 and links out to available PyTorch conda-forge builds.

• The CUDA version guidance and use of nvidia-smi to check the driver version are well-placed.
• The URL https://anaconda.org/channels/conda-forge/packages/pytorch/files?file_q=cuda uses a non-standard Anaconda path structure; the conventional package-files URL is https://anaconda.org/conda-forge/pytorch/files, and the file_q=cuda query parameter may not be supported.
• A blank line is missing between the closing code fence (line 29) and the "Choose a cuda-version…" paragraph (line 30), which is a minor but inconsistent style compared to the rest of the README.

Confidence Score: 4/5

• Documentation-only change; safe to merge with minor link verification recommended.
• The PR is a pure documentation update with no code changes. The core guidance is sound. The only meaningful concern is whether the linked Anaconda URL resolves to the intended page — if broken, it leaves users without an easy way to discover supported CUDA versions. Everything else is minor style.
• README.md — verify the linked PyTorch builds URL resolves correctly before merging.

Important Files Changed

Filename	Overview
README.md	Adds a note warning users about potential CPU-only PyTorch installs when the default conda solve picks cuda-version=13, with a workaround command and a link to available PyTorch builds. The link URL uses a non-standard Anaconda path that may not work as intended, and a blank line is missing between the closing code fence and the following paragraph.

_{Last reviewed commit: 6581e8e}

[scal444]

I don't think this is exactly right. cuda-version should

1. Be eq or less than your nvidia-smi
2. Have a supported pytorch version in conda-forge

Suggest linking the torch page here and having the instructions say pick a cuda-version <= nvidia-smi driver version that has a matching torch major version that they want + the right cuda version.

Alternatively. Suggest they install torch first. Check that it's GPU enabled. Then install nvmolkit

[evasnow1992]

Added two install options: (A) specify cuda-version explicitly — must be ≤ nvidia-smi version with a matching PyTorch build on conda-forge (linked the builds page), and (B) install PyTorch with GPU support first, verify with torch.cuda.is_available(), then install nvmolkit.

[scal444]

Isn't pytorch-cuda dead? It's from the pytorch channel which they no longer use. 12.4 was the last one they support. Current support is only through conda-forge

[evasnow1992]

Removed it and consolidated the two options into one since without pytorch-cuda, the "install torch first" approach will require specifying cuda-version explicitly, making it essentially the same flow as option A split into steps.