Evaluate by batch transform instead of endpoint

[athewsey]

The XGBoost notebook currently evaluates models by deploying to endpoint and then feeding data through (and the code to delete endpoints is right at the end).

It would be better to use batch transform jobs, since otherwise evaluating successive experiments could end up with multiple endpoints lying around (and maybe reaching quota once too many are created).

Would still be good to keep the endpoint deployment demo code around - perhaps at the end - as a reference