Add Rosenbrock-2S method

[BenWibking]

This adds the ROS2S integrator from https://doi.org/10.1016/j.apnum.2012.08.001.

It is a three-stage, second-order, stiffly-accurate, L-stable Rosenbrock-Wanner method with an embedded error estimator.

[BenWibking]

@zingale Here's the draft PR. This does not have all the performance optimizations that I tried.

[BenWibking]

This version wins by a few percent on CPU [Mac laptop] over VODE for primordial_chem when building both with Clang (which is much faster than gcc on CPU for these problems).

The performance delta is significantly larger on a Hopper GPU (in favor of Rosenbrock).

[zingale]

this gets an FPE if I build and run test_react as:

make INTEGRATOR_DIR=ROS2S DEBUG=TRUE
./main3d.gnu.DEBUG.ex inputs_aprox13  amrex.fpe_trap_invalid=1

[zingale]

it seems to be that it is generating really negative mass fractions and that's causing the failure, so we probably need more step rejection logic.