Load Neural Net

[jatkinson1000]

Closes #11

This is built on top of #9 which is itself build upon #4 and #5.

Testing using the CAM_GW_Net case on Derecho.

• Add filepath to neural net that can be read in from the namelist
• Read in and scatter to processes
• Generate a net on each process in the setup
• Destroy at the end of the physics routine

[jatkinson1000]

This is now ready for review.

It loads the net during the init phase of the simulation, and adds a gw_final subroutine to destroy it at the end.
The net is then available to be called within the gw_drag module as a module variable.

[yqsun91]

I have gone through this and it looks good to me.

One question? are we making additional changes to save the tendency term? utgw & vtgw for example. I did not find it here. But I think it is needed.

[jatkinson1000]

Thanks @yqsun91
Good question - Yes, we will save those variables, but that will be part of #14 - here I am specifically focussing on being able to load a net using FTorch and set up the surrounding infrastructure.

[jatkinson1000]

Note that due to delays in reviewing this PR went stale as it has been incorporated into #14 which builds on top of it.