[libcxx] Improve accuracy of complex asinh and acosh

Summary:
Currently std::asinh and std::acosh use std::pow to compute x^2. This
results in a significant error when computing e.g. asinh(i) or
acosh(-1).

This patch expresses x^2 directly via x.real() and x.imag(), like it
is done in libstdc++/glibc, and adds tests that checks the accuracy.

Reviewers: EricWF, mclow.lists

Reviewed By: mclow.lists

Subscribers: christof, cfe-commits

Differential Revision: https://reviews.llvm.org/D41629

llvm-svn: 325510