fix inference and use Base.precompile

[juliasloan25]

Currently, we run the coupling loop for 2 timesteps before running for the full simulation length and timing it. As part of #706, we tried to use Base.precompile instead, but this didn't increase the accuracy of the timing. This is probably because of type inference failures of solve_coupler!. We should fix the inference, then replace the initial two timesteps with Base.precompile