Implement a thin "data loading" layer to help ML training

[JackKelly]

The bulk of nowcasting_dataset is about saving pre-prepared batches of data to disk. But it feels like there's perhaps a need for a set of simple tools to help loading those pre-prepared batches into ML models during training. This has come up a few times in other issues:

• As mentioned in Use NamedTensors to name dimensions #25, perhaps write an xarray accessor function so we can go straight from xarray to pytorch named tensors.
• Train on subsets of the pre-prepared batch. e.g.
  • Option to change the length of history data to show to the model. e.g. always save 1 hour of history to disk. But then have a simple param to reduce that to, say, half an hour for some ML experiments (suggested by @jacobbieker in Add Rainfall Radar data #80)
  • Always save large satellite images (see Use 'large crop' of satellite image to provide a wider geographical context #87) with every example, but provide the option to pick just the centre crop during ML training. Or to down-sample the large image and provide native-res centre crop. (Also see https://github.com/openclimatefix/predict_pv_yield/issues/65)
• Maybe explore spatial reprojection on-the-fly (almost certainly too expensive; but might be worth a try). See Spatially reproject satellite data within nowcasting_dataset #92
• Standard way to load batches into a PyTorch DataLoader.
• Data augmentation (like flipping the image, see #78)
• Position encoding (e.g. converting the latitude & longitude of each PV system to an encoding that's ready for use in a fully-attentional model). See Encode "real world" positions of the input data perceiver-pytorch#9

Tasks like subsetting the data should be done as upstream as possible, so we only load from disk the data we want.

Tasks like data augmentation could perhaps be done in PyTorch 'transforms'?

[JackKelly]

This is implemented by nowcasting_dataloader so I'll close this issue for now (unless I've misunderstood?!)