Investigate the effect of different weightings for vertical levels

[sadamov]

The higher vertical levels show poor skill during inference/forecasting. This is a common issue and is often solved by using a weighting scheme during training. I suggest to downweight the highest layers. But of course one can also try the opposite and emphasize the models attention (weights) in these upper layers.

1. change the weights in the constants.py script, currently they are all set to one

# Vertical level weights
level_weights = {
    1: 1,
    5: 1,
    13: 1,
    22: 1,
    38: 1,
    41: 1,
    60: 1,
}

3. start a training with 40 epochs and validation every 20 epochs
4. compare loss curves to previous model runs on wandb