memory2 - per-stream metadata kv store

want to attach arbitrary key/value config to a stream and have it persist alongside observations.

first usage of this is record reply https://github.com/dimensionalOS/dimos/pull/1976 storing original topic names for streams (but I assume this is just first usage, embedding model name for embedding stream would be another example)

store.streams.lidar.metadata.calibration_offset = 3
store.streams.lidar.metadata.set("frame_id", "base_link")
store.streams.lidar.metadata.update({"sensor_model": "mid360", "rate_hz": 10})
store.streams.lidar.metadata["calibration_offset"]  # 3

distinct from per-observation Observation.tags (dimos/memory2/type/observation.py:57), which is immutable-derive (obs.tag(k=v) returns a new observation). keeping the APIs separate on purpose — observations are immutable, stream metadata is mutable config that's set once and read often.

requirements:

• attribute + dict-style get/set, plus set() / update() / delete()
• persisted by the backend (sqlite store gets a kv table; null/memory stores keep a dict)
• json-serializable values only at first; reject anything else loudly
• survives store reopen

should we...?

• make it Notifier-watchable so subscribers get told when metadata changes?
• expose it on the Backend directly so transforms/embedders can read it (e.g. a vectorstore reads embedding_model from stream metadata)?

related: #1652 (postgres store would also need this).

Synced from DIM-845 by summer

Start with dimos/memory2/type/observation.py:57 to understand how metadata differs from observation tags, then review the memory2 store implementations for sqlite, null, and memory persistence. Done means attribute and dict-style access plus set(), update(), and delete() accept only JSON-serializable values, persist across store reopen, and satisfy the open decisions about Notifier and Backend exposure.