Add macro to register CPU kernel for all arch types (#70332)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/70332
Idea to avoid recompilations: what if we introduce a new macro REGISTER_ALL_CPU_DISPATCH that registers the same kernel across all CPU arch types? We'd call this from native/Convolution*.cpp and wouldn't need to move any logic underneath the native/cpu dir. That would simplify these PRs quite a bit and would also avoid the recompilation. Wdyt about this approach?
Test Plan: Imported from OSS
Reviewed By: bdhirsh
Differential Revision: D33301403
Pulled By: jbschlosser
fbshipit-source-id: d7cc163d4fe23c35c93e512d1f0a8af8c9897933
