Fix consistency of precomputed quantities for Rosenbrock

[dennisYatunin]

Purpose

This PR fixes a bug in the Rosenbrock timestepper that results in inconsistent precomputed quantities in ClimaAtmos.

Since the first stage of an s-stage timestepping scheme is typically the same as the final state of the previous timestep, we can chose to update p based on each of the stages U_1--U_s, or we can choose to update it based on the stages U_2--U_s and the final state u. The Rosenbrock timestepper is currently using the first option, whereas the ARK methods use the second option. With the first option, the value of p between timesteps is not consistent with the value of u, since it did not get updated when u was computed and still has the value from U_s (for FSAL timesteppers, u = U_s, but SSPKnoth is not FSAL). The second option fixes this issue, so the Rosenbrock timestepper needs to be updated to use it.

Also, our computation for p modifies the state by updating the boundary face values of u_3. With the first option, only the intermediate stages get updated, and the final state u is left with unconstrained boundary face values of u_3. This has resulted in nonsensical surface velocities in ClimaAtmos simulations with topography (these velocities do not seem to be used for any tendencies or diagnostics, which is why they have gone unnoticed for some time).

---

• I have read and checked the items on the review checklist.