Due to how conversions were implemented previously, 1TB failed to
parse on 32-bit, as we were overflowing during computation, even
though the final result would fit easily.

This patch moves the parsing of the scaling factor to a separate
function, and all the conversions are done via the Rational type
(which has unlimited arbitrary precision), and conversion to the
desired type only happens at the last step.

The unit-tests are adjusted too, unfortunately they use the same
algorithm as the code… suggestions on how to improve things are
welcome.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>