Soft Prompt Tuning for Augmenting Dense Retrieval with Large Language Models