Improve convergence in v_from_i with large arguments

[cwhanse]

We currently use 3 iterations of a Newton method when the argument to the lambertw term is calculated in log-space. As the magnitude of the argument increases, the number of iterations should also increase to maintain the needed precision.