Estimation duration now uses the already parsed GCode instead of reparsing it.
It also computes a per layer duration estimation which can probably be used to
produce better ETAs.

The only difference is that it does not compute duration for Z layers changes,                                                                                                                        but it was probably super wrong already given how it is done (it only changes
the estimation by than 2s over 2 hours of print and 54 layers, and the feedrate
stuff mixes all the axes together).

I also detected a few potential issues in the code, which are marked by FIXMEs.