Extensive use of abs for complex tolerances

I just profiled `expintx(1,z::Complex)` and found the majority of the time is spent checking the convergence using `abs(stuff) < ϵ`. Would there be a problem with simply squaring all of these and using `abs2(stuff) < ϵ^2`? This is much much faster, especially when `stuff::Complex`.

Since `ϵ=10eps(T)`, I suppose that could lead to some issues. But I would much prefer to slightly reduce the tolerance than the current waste of compute.

I'm happy to do the PR if it's likely to get merged in.