Metal does not support floating point atomic addition on shared memory (block-local), but right now KernalAbstractions.jl would lower to invalid instructions, which prevents kernel from running.

I'm happy to make a PR, but is there a rope I can follow in terms to adding a different lowering for this?