AArch64:: Use 3 mov instructions to implement register swaps

[david-arm]

According to the software optimisation guide mov instructions between GPRs may be zero latency, which is even cheaper than using 3 eors.

[meta-codesync[bot]]

@facebook-github-bot has imported this pull request. If you are a Meta employee, you can view this in D86202186. (Because this pull request was imported automatically, there will not be any future comments.)