Use f64 instead of u64 in PascalRowIter to avoid integer overflow

[lars-frogner]

Something like this works up to n = 131 when F is f32 and up to n = 1020 when F is f64. Above these limits pascal exceeds the highest representable value and becomes infinity.

[ickk]

Thanks again Lars!

Regarding this approach, I'm not sure exactly what effect the inexactness of floats has on the quality of the output, but given we use this only for generating initial guesses it's probably a good compromise

Eventually I'll read through Dario Bini's paper (DOI:10.1007/BF02207694) #2 which I think replaces the procedure for generating initial guesses, and if we're lucky it may serve to further relax this constraint, but there's no rush. At the moment we expose the internal::aberth_raw function, so someone sufficiently motivated could supply their own initial guesses to work around the limitation if they needed

[lars-frogner]

Just to be sure, I compared the outputs of the f64 based implementation of the iterator with that of the u64 based implementation for all n < 64. They agree to within a relative error of 1e-15, so the inexactness shouldn't pose an issue as far as I can tell.