Training LLMs to be Better Text Embedders through Bidirectional Reconstruction