This PR makes the wind models work on GPUs.

The test was done with a wind farm with 49 turbines with the Fitch model.

1. Fig. 1 shows the contours of velocity deficit at t=1500 s from the CPU and GPU runs, showing they give identical results.
2. Fig. 2 shows the wall clock time per time step in seconds for 64 CPU cores (AMD EPYC 7763 (Milan)) using MPI, and a single GPU (NVIDIA A100) on Perlmutter for three different mesh sizes. The speed up factor for each mesh size is shown as well.