Large-Scale Representation Learning from Visually Grounded Untranscribed Speech