Let user pass train/val/test data split

[radka-j]

Currently we do a random train/val/test split. There are use cases where the user might have a more challenging test split in mind (e.g., to better evaluate out of distribution model performance). We should support this.

[sgreenbury]

An option for this would be to extend the AutoEmulate init API such as:

class AutoEmulate:
    ...
    def __init__(
        self,
        ...,
        validation_data: tuple[InputLike, InputLike] | None = None,
        test_data: tuple[InputLike, InputLike] | None = None,
        ...
    ): ...

where the tuples are (x, y) pairs for the validation and test respectively.

[radka-j]

@sgreenbury yes that's what I was imaging. And we do the split as currently implemented in case these are not passed.