Two major reasons:

* This avoids the public dependency on `num_ratio`.
* The internal representation can be changed. This already proved
  necessary once (from u16 rational) and in this manner we are free to
  do it again by adding an inner enum representation.

Note that we already have rounding semantics for the meaning of Delay as
gif rounds to the nearest 10ms when writing frames. Any more or less
precise representation can be made to round to the closest u32 ratio.

With the original available as a millisecond ratio this can easily be
done to the accuracy of nanoseconds, which is the finest available.

Given the prior usage, it does not make much sense to provide an exact
conversion. However, it should be clear that precision is not finite and
the representation is not exactly that of a Duration.

The implementation turned out to be fairly involved as the num-fraction
crate does not provide approximation algorithms for fractions. No fear,
we code them ourselves with extensive testing to cover much more than
the required accuracy tests.