Use App::Prove::State to store the timings for the tests, and if

Use App::Prove::State to store the timings for the tests, and if
timings are available, reorder the parallelisable tests to run the
slowest first. Timings for a second run are 18 seconds less for me:
Files=1553, Tests=209393, 459 wallclock secs (94.89 usr 13.16 sys + 638.19 cusr 58.59 csys = 804.83 CPU)
Files=1553, Tests=209393, 441 wallclock secs (82.83 usr 13.90 sys + 622.13 cusr 59.20 csys = 778.06 CPU)

p4raw-id: //depot/perl@34192