Previously, the solver would recompute the Jacobian at every iteration. But, if
an iteration fails, the current point is not moved (only the radius of the
trust region changes), hence it is not necessary to recompute the Jacobian in
that case. This commit implements this optimization.