Don't throw in eigs on semidefinite Bs for generalized eigenproblems.

[andreasnoack]

@madeleineudell With this it should be possible to handle semidefinite Bs via explicit shift. See the added test for an example.

[madeleineudell]

@andreasnoack , this is fantastic. thanks!

[tkelman]

similarly, be very careful about your minimum julia version dependency if you rely on this in 0.4.7

[andreasnoack]

I think we might want to revert this. The behavior reported in JuliaLang/LinearAlgebra.jl#488 is simply too problematic and it is pretty standard to assume that B is PD in the symmetric generalized eigenvalue problem. That is what LAPACK does. @madeleineudell On the mailing list, you mentioned that B was often not positive definite. I assume that this might be the case for optimization problems. Could you explain the properties of A and B for typical and maybe also some less typical optimization problems. Maybe we can find a solution that can work you without opening the door for silent errors like the ones reported in JuliaLang/LinearAlgebra.jl#488.