Use LAPack's `zhseqr` / `dhseqr`?

LAPack's Hessenberg eigenvalue routines use half the memory. I'd be willing to make a pull request that moves (and modernizes) ApproxFun.jl's implementation over: https://github.com/JuliaApproximation/ApproxFun.jl/blob/master/src/LinearAlgebra/hesseneigs.jl

Note that there is an unfortunate side effect that you don't get auto-column scaling, so this may have side effects. I got around this with an adhoc column scaling.