Change config example to call validation split val instead of validation

[leifdenby]

This isn't technically a bug because mllam-data-prep doesn't try to enforce any kind of specification wrt what the outputs are called, but the new datastores functionality in neural-lam mllam/neural-lam#66 assumes that the validation split is called val rather than validation, so it would be nicer to give the examples in mllam-data-prep to be consistent with that