embeddings: improve model management UX

[totocaster]

Summary

README highlights a need for better embedding model management (preset docs, cache overrides, richer download progress). Track this work so we can plan UX improvements.

The scope of work here is a little bit bigger than described below, as I'm also considering making embeddings completely custom for a user-supplied interface. Perhaps embeddings should become a plugin for the arrowhead.

Tasks

• Document each embedding preset (quality, performance, memory footprint).
• Allow overriding cache directories and reusing downloaded models across vaults.
• Surface richer download progress / status in CLI and daemon logs.

Definition of Done

• Users can understand and configure embedding presets from docs and CLI help.
• Cache location overrides are persisted and tested.
• Download progress is visible without enabling verbose logging.