Support dynamic loading of shared NVRTC libraries

[c0d1f1ed]

Warp currently statically links NVRTC (plus nvrtc-builtins, nvptxcompiler, and optionally nvJitLink) into its native library. To support testing of pre-release versions and development builds of NVRTC we should support letting the user point Warp at a directory of shared NVRTC libraries that override the statically linked ones at runtime.

Design: dynamic-nvrtc.md

[shi-eric]

I think https://github.com/conda-forge/warp-lang-feedstock also has some patches they use to allow Warp to use dynamic CUDA libraries. It would be good if the solution was also compatible with the CUDA components that are available on conda-forge so they can drop the need for a separately maintained patch.

Also, is it not possible to attach the design file inline? It's not very convenient to have to download the markdown file locally to see what's inside.