`PushN` peephole opts need a separate pass

[ike709]

In short, the various PushN opts like PushNFloats need to be moved to their own optimization pass after the other peephole opts have had a chance to run.

They currently interfere with other optimizations like const folding, which check for PushFloat followed by PushFloat.

Further explanations and why alternate solutions are suboptimal can be found on Discord. I can't be bothered to dig them up since I intend to fix this after #2054 is merged anyways.