Fix denominator of roundto result

Multiprecision rational created from a double value may have large power
of 2 denominator since fractional decimal numbers can not be represented
as binary floating point numbers. It leads to failed assertion when
result is compared to a value converted directly from strings.
Use integer multiprecision arithmetics to round numbers to ensure
proper denominator. Inspired by python gmpy2 package
<https://github.com/aleaxit/gmpy/blob/3e4564ae9d/src/gmpy2_mpq_misc.c#L315>

The change makes `roundto` symmetric for positive/negative arguments.

Halves are rounded to nearest even. Rounded away from zero are discussed
in #1663 and it may be achieved with minimal modification.

- See #2329
- Closes #1983

0 files changed, 0 insertions, 0 deletions