C3PO: Learning to Achieve Arbitrary Goals via Massively Entropic Pretraining