Enhancement to Defining Train and Validation Splits docs

[robmarkcole]

In the main readme, under Defining Train and Validation Splits two approaches are listed for splitting the dataset, however only the second (tags) approach as an example.

I attempted approach one (Assign windows to different groups) using a script which copied data from windows/global_soil_moisture/... into a format windows/train|val|test on a completely different drive

I then assumed I could config:

data:
  class_path: rslearn.train.data_module.RslearnDataModule
  init_args:
    path: "/my-new-drive-location" # contains windows/train|val|test
   ...
    train_config:
      groups: ["train"]
      transforms: []
    val_config:
      groups: ["val"]
    test_config:
      groups: ["test"]

But this results in a long and confusing error from the datamodule. The root cause is that each window metadata.json hard codes the original group (global_soil_moisture). I addressed this with a code change on my branch (below). I think an example of approach one would prevent others from taking the path I did, or the approach above could be supported as many datasets come in this format
Thanks

# In ModelDataset
    def _serialize_item(self, example: Window) -> dict[str, Any]:
        data = example.get_metadata()
        # Persist the actual on-disk location because some datasets relocate windows
        # without updating the group stored in metadata. Using a relative path keeps
        # backward compatibility with existing config files while allowing us to
        # reconstruct the window even if the metadata `group` no longer matches the
        # directory hierarchy (e.g. splits moved from a single group into
        # train/val/test folders).
        try:
            relative_path = example.path.relative_to(self.dataset.path)
        except ValueError:
            # Path may fall outside the dataset root (e.g. remote URLs); fall back to
            # storing the absolute path in that case.
            relative_path = example.path
        data["relative_path"] = str(relative_path)
        return data

    def _deserialize_item(self, d: dict[str, Any]) -> Window:
        if "relative_path" in d:
            window_path = self.dataset.path / d["relative_path"]
        else:
            window_path = Window.get_window_root(
                self.dataset.path, d["group"], d["name"]
            )
        return Window.from_metadata(window_path, d)