Multimodal Fine-grained Context Interaction Graph Modeling for Conversational Speech Synthesis