Adapting LLMs for Efficient Context Processing through Soft Prompt Compression