Prompt Tuning with Soft Context Sharing for Vision-Language Models