compiler_rt: update memcpy to compare usizes at a time

[nektro]

cc @mikdusan noticed this improving stage2 compilation performance of stage3 by slightly more than 2x on macOS when using our own implementation

libSystem.memcpy: 299.05 seconds memcpy.zig: 996.01 seconds memcpy_usize.zig: 443.22 seconds

[nektro]

further notes:

• memmove can be modified to call this function
• memcpy can be modified to use this same optimization

[Jarred-Sumner]

Couldn't this use SIMD to go in chunks of 16/32 and copy that way? Should be straightforward with @Vector

[rohlem]

Uhm, both the title and the code comment mention "comparing" usizes - that should read "copying", right? Did the first comment also mean to mention memcmp instead of memcpy? Although 4 people have already looked at this, so maybe I'm bugging.

[nektro]

no @rohlem you're right