Simplify Moves a bit.

1. Change all Move overlap checks to a more obvious and potentially more efficient (1 conditional jump to .Lback instead of 1 more to .Lforward) count > unsigned(dest - src).

2. Write last two vectors after 2× loops unconditionally instead of an extra check. This is A BIT DUBIOUS (removes another jump, but by always performing the read+write it was jumping over, with 0.37% more chance for a cross-page write), but is shorter.

3–4. Also includes parts of !551 not related to ERMS; I’ll remove them from !551...