Variants of Rana PC

The original Rana-type preconditioners are based on finding a triangular approximation to the Butcher matrix with LDU factorization.  However, papers like https://arxiv.org/abs/2206.08991 suggest other methods like triangular L that minimizes || I - L^{-1} A ||_2 or kappa_2(L^{-1} A).  These should be easy to implement with numpy/scipy?
