[llvm] AArch64: Optimize memmove for non-power-of-two sizes (PR #168633)

[Nashe Mncube via llvm-commits]

https://github.com/nasherm requested changes to this pull request.

I like the idea of this patch. But I have some comments

The codegen from this patch doesn't seem that much better for testcases ```move33```, ```move49```, and ```move65``` https://godbolt.org/z/nEYvE5YY7

https://github.com/llvm/llvm-project/pull/168633