This posed an issue when the type parameter was a typedef.

Implemented Smith's method, which avoid overflow for all but
the most extreme exponents. This agrees with what is used in NumPy.

avoid useless formatting operations when we know at compile time that the result is a Unicode string

implement somewhat useless but obvious and straight forward string conversion of the C error return type

replace dead left-over code by the actually intended fast bint->True/False formatting implementation