Update CUDA chol() implementation to new cuSolver API, and other cleanups

A couple cleanups here:

• Adapt cuda::chol() and opencl::chol() to not throw exceptions, but instead return a std::tuple<bool, std::string>, like svd(), so that exceptions can be suppressed for some versions of chol() (like Armadillo).
• Use cuSolverDnXpotrf() instead of deprecated type-specific versions.
• Handle some memory leaks that could happen on failed cuSolver calls.
• Check info code to ensure decomposition was successful, and print reasonable error messages if not.
• Adapt tests to use TEMPLATE_TEST_CASE().