Record loss by time of day/day of year

[jacobbieker]

We want to know how out model's loss changes throughout the year and time of day. Do they do worse in the morning, when the past satellite data is darker and maybe less informative? Do they do better in the summer or winter? This could help with choosing how many examples we need to show the model from different times and seasons to ensure it works well across the year.

As nowcasting-dataset gives the date of the input sequence as part of the batch, that should be enough to start recording the loss for these time periods.

Moved from openclimatefix/satflow#69