feat(multimodal): build multimodal capabilities for ReMe

**Background**

ReMe currently focuses primarily on text-based memory. To support agents that interact with images, audio, video, and mixed-content documents, ReMe needs a unified multimodal memory capability that preserves its local-first and file-native principles. Multimodal source files should remain user-owned and authoritative, while derived representations such as embeddings, indexes, captions, and metadata remain rebuildable.

**Changes**
- Define extensible schemas and interfaces for multimodal memory content and metadata
- Support ingesting and storing images, audio, video, and mixed-content documents without replacing the original user-owned files
- Add pluggable preprocessing and embedding pipelines for different modalities
- Enable multimodal indexing, retrieval, and cross-modal search
- Expose multimodal capabilities through the existing service, client, and job interfaces
- Add configuration, documentation, and focused tests for the new workflows