Stable psd solve

[calebweinreb]

More stable version of psd_solve that addresses #334.

• The input matrix is symmetrized
• Only the diagonal is boosted
• Uses cho_factor / cho_solve instead of linalg.solve

[slinderman]

Thanks! I believe @gileshd had originally implemented something similar, but we hit numerical stability issues in our tests. Yours may differ in that it explicitly symmetrizes as well as boosting the diagonal. I'll go ahead and merge!