Implement 1x1 nonlocal GW model into CAM

[TomMelt]

fixes DataWaveProject/nonlocal_gwfluxes#48
fixes DataWaveProject/nonlocal_gwfluxes#51
fixes DataWaveProject/nonlocal_gwfluxes#52

Add 1x1 CNN non-local gravity wave model into CAM

This is an initial implementation of coupling the nonlocal gw models into CAM.

• Instructions on how to run are included in the README.md
• Path to traced/scripted NN can be provided via gw_nlgw_model_path namelist var
• nlgw model can be enabled by setting use_gw_nlgw=.true. in the namelist
• model std dev.s and means are currently hardcoded (this could be changed in a future PR)
• era5 ak and bk coefficients are also hardcoded to interpolate between the ERA5 pressure grid and the CAM pressure grid
• currently inference takes place on the GPU and as such this requires a GPU TorchScript model

Note

If you follow the instructions in the README.md before this is merged you need to checkout branch nonlocal-gws instead of datawave_ml

TODO

• output tendencies to fix output utgw vtgw via CAM outputs nonlocal_gwfluxes#52
• verify model works with latest trained model (epoch 85) to fix use latest version of weights (epoch 85) nonlocal_gwfluxes#51

Merge Dependencies First

• Fix flux derivative #34
• Add tendencies and fluxes to variable output #30

[TomMelt]

@amangupta2 can you double check the units of the inputs/outputs vs training data and expected model output

[amangupta2]

Sorry for my late response on this, but all the units seem consistent.

[amangupta2]

I have scripted the model but have been unable to run CAM successfully, so perhaps I can merge this PR tomorrow morning after our discussion?

[TomMelt]

Once we have merged #30 and #34 I think we can merge this. @ma595 would be good to have one of the longer runs to confirm results look sensible