This commit started off because I wanted to remove MPFR as a dependency.
After an initially-incorrect attempt at doing so, I discovered two more
problems. First, the test suite didn't have any tests verifying that MPFR
actually fixed the problem it was introduced to fix (correctly rounding
results like 0.1*1000 to 100 instead of 99). Second, MPFR gets 0.1*1000,
but still makes rounding mistakes in other cases--I found this out by
fuzzing after my initial incorrect patch.

So now I've replaced MPFR with exact rational math in libgmp, and added
regression tests for the rounding errors in string parsing. To do this, I
had to do the parsing manually; I followed the same regex approach that
the existing function uses.

Signed-off-by: Jon Oster <jon.i.oster@gmail.com>

Switching to internal parsing logic caused two changes to existing behaviour.
"1. KiB" was previously accepted, and parsed the same as "1.0 KiB". And "e+0"
was previously not accepted, raising an error; with the new parsing it's parsed
as 0. This was found by an external test suite, so we're adding these cases to
the internal suite as well.

Signed-off-by: Jon Oster <jon.i.oster@gmail.com>

Allows parsing of strings with a decimal but no decimal part, e.g. "1. KB".
Fails to parse if there is no number present by adding a validation check
to ensure that either the integer part or the fractional part has len > 0.

Signed-off-by: Jon Oster <jon.i.oster@gmail.com>