high-res realistic setup is unstable

[Thanduriel]

I have been running the setup from the integration test (i.e. northern hemisphere, TOPAZ4 + ERA5 forcing) for the different available grids. It works fine up to 6.25 km, but the 3.125 km grid becomes unstable after just a few steps, even for very small time-steps (e.g. 2s). I can't pinpoint a specific point where the instability originates but it does look like it expands from the coast-line (showing hice steps 3 and 4):

The setup was created using the scripts era5_topaz4_maker.py and make_initNH.pyy with the only modification being an added single cell boundary around the domain. The whole setup is available here and can be run with the current develop version. I also tried various tweaks to pin down the problem including a landmask consisting of just the outer boundary and dummy thermodynamics but it seems to persist.