MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference