Fix rounding errors in time conversion

Fix rounding errors in time conversion

* src/timefns.c (frac_to_double): Pass FLT_RADIX to mpz_sizeinbase
instead of doing the radix calculation ourselves, not always
correctly.  Fix off-by-one error in scale, which caused
double-rounding.
(decode_time_components): Use frac_to_double (via decode_ticks_hz)
to fix double-rounding error that can occur even though
intermediate results are long double.
* test/src/timefns-tests.el (float-time-precision):
Test the above fixes.