Support negative magnitudes

[chiphogg]

We add a new base, au::Negative. This always goes at the beginning of
the pack. We define rational powers such that Negative disappears
when the numerator (expressed in lowest terms) is even; otherwise, it
remains if the denominator is odd, and we produce a hard compiler error
if the denominator is even. For products, the usual machinery works in
most cases, but if both inputs have Negative, then we remove them
both, and delegate to the product of what remains.

This brings new magnitude traits: Abs for the absolute value, and
IsPositive to tell whether a magnitude is positive: formerly, these
would have been the identity, and trivially true, respectively.

For obtaining magnitude representations, we add a new error state
outcome, ERR_NEGATIVE_NUMBER_IN_UNSIGNED_TYPE. If this doesn't
happen, we simply delegate to computing the magnitude of the absolute
value, and negate the result. This does mean we won't produce an answer
for the minimum value of signed integral types, but we can complete that
later.

Finally, we support magnitude labels by prepending a - sign. We
always prepend this sign, even for unlabeled magnitudes, because it's
important to convey sign information.

Note that this change enables users to create broken units and
quantities. If they use negative magnitudes with units, it will not
produce correct results. But given that (a) nobody has ever been able
to form these magnitudes before, (b) the functionality is undocumented,
and (c) we will follow up very soon with changes that restore this
functionality, this does not seem like a serious concern.

Helps #370.

[hoffbrinkle]

Should we update the comment?

[chiphogg]

Done.