Fuse more kernels in SSPKnoth

[charleskawczynski]

I realized that now we default to SSPKnoth in ClimaAtmos, one reason why fieldvector kernels take up so much time in CliMA/ClimaCore.jl#2067 (comment), is that we are not sufficiently fusing kernels in SSPKnoth.

[charleskawczynski]

i.e., step_u!(int, cache::RosenbrockCache{Nstages}) where {Nstages}

[charleskawczynski]

Quantitatively, this build shows that the 4 gpu dry baro wave spends 27% of its time in fieldvector operations, which are mostly done in the timestepper:

[ Info: Statistics across 299016 total kernels
                                  Kernel duration percentage
                          ┌                                        ┐
                     fill ┤ 0.0794944                               
       dss_load_from_recv ┤ 0.263896                                
     dss_fill_send_buffer ┤ 0.274443                                
                dss_ghost ┤ 0.369122                                
          dss_local_ghost ┤ 0.453303                                
              CUDA memcpy ┤■ 1.05338                                
              CUDA memset ┤■■ 1.5149                                
                dss_local ┤■■ 1.76417                               
          dss_untransform ┤■■■ 2.66246                              
            dss_transform ┤■■■ 2.71636                              
                   copyto ┤■■■■■■ 6.05631                           
single_field_solve_kernel ┤■■■■■■■■ 8.18527                         
                 spectral ┤■■■■■■■■■■■■■■■■■■ 17.5483               
          CuKernelContext ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 27.1507     
                  stencil ┤■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■ 29.9078  
                          └                                        ┘