cmd/compile: improve loop rotation

[josharian]

This is a follow-on to #20355 for the 1.10 cycle. See CL 43491 for more discussion of these points.

• Should we work from innermost loops out?
• Do we want to move contained loops?
• Do we want to reinstate some form of loop skipping?
• Other improvements?

See also #18977 and #20356.

cc @randall77 @dr2chase @cherrymui