Prior to this, Duration simply derived Debug. Since Duration doesn't
implement `Display`, the only way to inspect its value is to use
`Debug`. Unfortunately, the derived `Debug` impl is far from optimal
for humans. In many cases, Durations are used for some quick'n'dirty
benchmarking (or in general: measuring the time of some code). Correctly
understanding the output of Duration's Debug impl is not easy (e.g.
is "{ secs: 0, nanos: 968360102 }" or "{ secs: 0, nanos 98507324 }"
shorter?).

This commit replaces the derived impl with a manual one. It prints
the duration as seconds (i.e. "3.1803s") if the duration is longer than
a second, otherwise it prints it in either ms, µs or ns (depending on
the duration's length). This already helps readability a lot and it
never omits information that is stored.

This `Debug` impl does *not* respect the following formatting parameters:

- fill/align/padding: difficult to implement, probably not worth it
- alternate # flag: not clear what this should do

Rounding is done like for printing floating point numbers. If the
first digit which isn't printed (due to the precision parameter) is
larger than '4', the number is rounded up.

Previously, the code would panic for high precision values. Now it
has the same behavior as printing normal floating point values: if
a high precision is specified, '0's are added.