Tests fail when compiling with "/fp:fast" with MSVC.

Checks when running "test-conversions", "test-ieee" and "test-strtod" fail, which is pretty much expected, since the fast floating-point behavior is deliberately imprecise.

See documentation here: https://msdn.microsoft.com/en-us/library/e7s85ffb.aspx

Should the library force "/fp:precise" behavior using the following pragma? Or shall it be up to the user to choose the correct floating-point behavior? In that case it would be good to add this to the documentation.

```
#pragma float_control(precise, on, push)
...
#pragma float_control(pop)
```