bug#26162: time-duration screws up negative durations

[Zefram <zefram@fysh.org>]

Computing a difference between two SRFI-19 times, using time-difference,
produces sensible results if the result is positive, but often nonsense
if it's negative:

scheme@(guile-user)> (use-modules (srfi srfi-19))
scheme@(guile-user)> (time-difference (make-time time-tai 0 1) (make-time time-tai 1000 0))
$1 = #<time type: time-duration nanosecond: 999999000 second: 0>
scheme@(guile-user)> (time-difference (make-time time-tai 1000 0) (make-time time-tai 0 1))
$2 = #<time type: time-duration nanosecond: 1000001000 second: 0>

The above is computing the same interval both ways round.  The first time
is correct, but the second is obviously not the negative of the first.
The correct result for the second would be

    #<time type: time-duration nanosecond: 1000 second: -1>

or possibly, at a stretch,

    #<time type: time-duration nanosecond: -999999000 second: 0>

(SRFI-19 isn't clear about which way it's meant to be normalised.
Having the nanoseconds field always non-negative is less surprising and
easier to maintain through computation.)

-zefram