Add other convergence criterion for kdyn=3 when the 2-norm of the residual is very small.

[JFLemieux73]

In idealized test cases, the 2-norm of the residual can be very small. With the current convergence criterion based on a drop of the residual (2-norm / initial 2-norm) the solver can keep iterating because the 2-norm does further decrease as it is limited by the machine precision. The solver stops iterating when the max number of nonlinear iteration is reach. @phil-blain FYI.