This is an attempt to fix #360. The algorithm produces
decimal representations which are guaranteed to roundtrip
and in ~99.8% actually produces the shortest possible
representation. So this is a nice compromise between using
a precision of digits10 and max_digits10.

Note 1:

The implementation only works for IEEE single/double precision
numbers. So the old implementation is kept for compatibility
with non-IEEE implementations and 'long double'.

Note 2:

If number_float_t is 'float', not all serialized numbers can
be recovered using strtod (strtof works, though). (There is
exactly one such number and the result is off by 1 ulp.)
This can be avoided by changing the implementation (the fix
is trivial), but then the resulting decimal numbers are not
exactly short.