Reimplement IPSet.Ranges with a (somewhat) easier to follow algorithm.

[danderson]

    The main goal was to fix #125 by providing a less tricky implementation
    of Ranges as a comparison baseline... But this implementation turns out
    to also be a bit faster and more memory efficient.
    
    Fixes #125.
    
    name         old time/op    new time/op    delta
    IPSetFuzz-8    16.3µs ± 3%    14.1µs ± 1%  -12.99%  (p=0.008 n=5+5)
    
    name         old alloc/op   new alloc/op   delta
    IPSetFuzz-8    2.60kB ± 0%    1.95kB ± 0%  -25.24%  (p=0.008 n=5+5)
    
    name         old allocs/op  new allocs/op  delta
    IPSetFuzz-8      33.0 ± 0%      30.0 ± 0%   -9.09%  (p=0.008 n=5+5)
    
    Signed-off-by: David Anderson <[email protected]>

[danderson]

Ready for review @bradfitz @mdlayher. Tests pass at all commits (I only add the previously-failing test case after the major commit that fixes the bug), so can be rebase-and-merged.

[mdlayher]

LGTM, and I appreciate the ASCII diagrams!

[bradfitz]

Nice!

[danderson]

Looks much better with the symbolic names for each case, thanks. Applied all feedback, going to submit and we can bikeshed on the exact names of helpers in a followup (I'm not super happy with some of the subtler ones).