TRACE: Grounding Time Series in Context for Multimodal Embedding and Retrieval