For relative mode, not only needs the retractions created to be actually relative, but it also needs to compensate on the next G1 (as opposed to absolute mode, where you'd want to go to the same absolute E position after). Rather than massively complicating the already gnarly code (once it found a G1 retraction, it scanned forwards to find G0 statements (which it rewrote to G1), so it would go over those _again_ in the middle (layer) loop). While this worked for absolute mode, but would be a nightmare to make work for relative mode as-is (if only because the compensation could _also_ potentially involve keeping track of things over the outer loop). As a bonus I think the resulting code is actually easier to read.
part of CURA-10092 -- should fix#14100
Otherwise if a retracted move had a different Z coordinate, this would not be put in the final print. It's pretty simple to extend it this way.
Fixes#7458.
