[META] Improve Helios production model

[JackKelly]

• Add back in the GSP timestep to the GSP query generator for satellite transformer #96
• More NWP variables #145
• Plot NWPs #164
• Try giving the NWP as multiple query elements. e.g. each variable is its own element (with a learnt channel ID) #163
• NWP query needs solar position, and step. #165
• Use the same order of features for each query. #166
• Instead of using a one-hot encoding of NWP step, it might be better to use nwp_init_time_utc_fourier where the other queries have time_fourier_t0.
• Tweak plot of national pv: Include date & y-axis label & title. #171
• Weight the first hour more highly #147
• All NWP variables
• More Gaussians #172
• More heads & higher D_MODEL #173
• train on more data. Enable the data loaders to take a list of periods to load, so we can validate on, say, every 5th week of 2021. #176
• Give the model the absolute PV capacity
• Weight the GSP loss by the PV capacity of the GSP?
• GSP ID of pixels & PV systems
• Predict multiple GSPs at once  #175
• Use hydra to specify config options #170
• Increase history #168
• Reduce weight of PV loss
• Increase learning rate?
• Compute nanmean and nanstd for all NWP variables.
• Learnable ID for Satellite, PV and GSP queries.
• Pad with learnt vectors, not zeros.
• Separate NWP query elements for different locations
• Downsample NWP using little CNN
• Try removing the RNN again.
• Don't reshape_time_as_batch in GSPQueryGenerator! #162
• Don't include GSP history in the query for time_transformer. We're already giving the GSP history into the SatelliteTransformer. And it may be confusing for the model to have the GSP history go to zero for the forecast horizons.
• Don't give the GSP history in any of the main queries. Instead give a separate input to the time_transformer which includes the GSP history, and a marker to say "this is history".
• Plot attention matrix
• Don't give PV history in any of the queries. It's already in the RNN.
• Write prepared batches to disk as a separate script #167
• Try using an RNN for GSP power #102
• Reduce duplication between PVQueryGenerator and GSPQueryGenerator #144
• Add noise to "updated" GSP PV_Live so it looks more like intra-day PV_Live, and train model #120
• Continue training the U-Net #156
• Train on PVoutput.org #118 (although first wait to see if Peter gets Passiv data into production)
• If the model is too big to run in production, then convert to Torch Script