This sounds a lot like the kind of stutter Dolphin's dynamic recompiler used to suffer when it first discovered a new shader.
IIRC, for faster boot times, Emacs binaries burn in some kind of low-level memory-image that gets dumped into an mmap region and then "revivified." I wonder if that's what Rosetta is choking on.
If so, I wonder whether, until an actual AArch64 build is available, it'd be possible to work around this problem by just rebuilding Emacs (for x86!) without that optimization enabled.