The method is rather naive - digits read as an integer, then multipled
by a power of 10 or 2. This does not always give exact results, but it's
close enough for now. A stub support for long double larger than 64 bits
is provided.
The presumed bug where the value computed without the sign overflows
even though the negative result can be represented is not actually a
problem, because this only happens with signed results and the temporary
value is computed as unsigned (thus with extra range).
This is enough to support the standard and likely the C++ library and
external programs to port, but also the most we can do without a proper
locale data storage and more target-specific developments that aren't a
priority right now.
