this changes llvm-reduce's phase ordering to run the delta passes that remove function bodies first. on a collection of real 77 IR modules that I'm using to test reduction speed/power, this reduces llvm-reduce's execution time by 20%.
the rationale for this change is simple: a delta debugger wins when it eliminates as much junk as possible very early. the passes that this patch puts at the front of the phase ordering are the ones that accomplish this.