Training: experiment with number of layers/channels++ @PPI

[inakbk]

We should figure out what the limit of PPI training is. Increase:

• num_layers (we have used 8, could be 16 or more?)
• num_channels (we have used 80, but could be 1080 or more?)

Other things to learn more about and see if we should change is:

• limit batches
• trainable_parameters
• max epochs
• iterations

Would also be an idea to do this with only surface variables like:

• salinity_1
• temperature_1
• u_eastward_1
• v_northward_1
• w_1
• zeta

but have not managed yet to get a training to work based on selecting the names of the variables in above list. See #28