Handle some edge cases better

[Harmen Stoppels]

• All zeros matrix gives zero norm of Hessenberg matrix: < should be <= in convergence criterion
• Division by zero when computing eigenvectors of quasi upper triangular matrix
• Sylvester equation can be very ill-conditioned with close eigenvalues -- currently it checks if the eqn is exactly singular, but we should probably improve the situation where it's close to singular