Dwell in the Beginning: How Language Models Embed Long Documents for Dense Retrieval