numerical accuracy of the `hermiteh` function

[barton-willis]

I didn't find documentation about the numerical accuracy of the hermiteh function, so the following example isn't a bug, but it might surprise some users:

`
julia> using ClassicalOrthogonalPolynomials

julia> n,x = 50, -9.182406958129317
(50, -9.182406958129317)

julia> function bigrational(r::Rational)
return BigInt(numerator(r)) // BigInt(denominator(r))
end
bigrational (generic function with 1 method)

julia> Q = hermiteh(n,x)
-1.7132524505571502e43

julia> H = convert(Float64, hermiteh(n, bigrational(convert(Rational,x))))
-1.917927837463344e43

julia> (Q-H)/H
-0.10671693841041378

`

If all has gone well, H is the nearest binary64 to the value of the 50-th degree Hermite polynomial using the exact rational value of the binary64 number -9.182406958129317. The relative difference between H and Q rounds up to 11%

[barton-willis]

Another example--this one has degree five:

julia> using ClassicalOrthogonalPolynomials

julia> bigrational(x::Rational) = convert(BigInt, numerator(x))//convert(BigInt, denominator(x));

julia> n,x = 5, 2.0201828704560856
(5, 2.0201828704560856)

julia> Q = hermiteh(n,x)
-1.4086460592372857e-14

julia> H = convert(Float64, hermiteh(n, bigrational(convert(Rational,x))))
-4.5687075073857527e-14

julia> (H-Q)/Q
2.2433324733535183

[dlfivefifty]

It's not really possible to have relative accuracy near roots with floating point arithmetic.