Evaluate Nix-built binaries for performance with different tuning

[ConnorBaker]

Does specifying hostPlatform.gcc.arch when importing Nixpkgs yield binaries which are more/less performant?

Some packages are smart enough to compile code for features not exposed by the host CPU (like OpenSSL and accelerated hash functions) and use runtime detection/dispatch, but not all do this. (Aside: do we enable GCC’s function multiversioning?)

Investigate the closures of some of the following to test for performance differences:

• nix
• OpenCV (are image registration routines compiled with support for avx256/512, for example?)
• PyTorch (would non-accelerator inference be faster?)