Mixed-modality Representation Learning and Pre-training for Joint Table-and-Text Retrieval in OpenQA