Just run Arch in a VM on the mac, using Apple's virtualization framework [1]. (No need for other wrappers; the framework is often easier to use directly.) I find no significant performance issues; a lot of software actually runs faster in a Linux VM than when I compile and run it on the mac.
The limitations are that the Linux VM's don't support save/restore of state (i.e., you have to shut down to stop the VM) and the graphics and device support is limited.
You might get more performance if you compile your own Arch distribution with all the flags necessary to enjoy modern CPU features in M-series (particularly M4+ with SME extensions). With M4-Pro memory should run 273GB/second for JAX and PyTorch.
However, AFAICT, PyTorch supports vectorization on Apple silicon by delegating to Apple's Metal API's, and that wouldn't be available from the Linux VM, so you might prefer running those on macOS directly.
To me the biggest draw of the M4 processor is that it supports CPU tracing (at least for Xcode-run code): actual, direct (not sampled) data about CPU internals like branching, cache misses, etc. If you're into performance, this will take you from working with a black box to seeing exactly what's up.
[1] https://developer.apple.com/documentation/Virtualization/run...