Wire memory retrieval through resource scheduler

[jaylfc]

Embedding and rerank calls currently talk to rkllama directly via qmd-client. Refactor so the memory_search skill and any RAG path submit Tasks to the scheduler with:

• capability = Capability.EMBEDDING
• preferred_resources = [npu-rk3588 with max_wait_ms=200, cpu-inference]
• priority = Priority.INTERACTIVE_AGENT

This is the load-bearing example in docs/design/resource-scheduler.md §Worked example — chat-while-generating. When image gen holds the NPU, memory lookups should transparently route to CPU embedding in ~500ms rather than blocking behind 34s of SD.

[jaylfc]

Scoping note after today's investigation: there is no direct embedding call on the controller side that can be wrapped in scheduler.submit for Phase 1.

• User memory search (user_memory.search) is SQLite FTS only — no embeddings involved
• Agent memory / QMD embed happens inside the per-agent LXC container via the container's own rkllama routing; the controller proxies text through to qmd_url and does not see the embedding call
• The qmd_client.embed() helper is defined but not called from any in-controller path

This becomes a real target in two scenarios:

1. We adopt qmd serve on the controller as a shared embedding broker (see feat: remote model server (qmd serve) for shared inference across clients tobi/qmd#511) — then agent QMDs talk to the controller's qmd serve and the controller's embedding call is in-process and wrappable.
2. We add an in-controller embedding endpoint to the user memory feature (semantic search instead of FTS).

Leaving the issue open but relabelling as blocked on Phase 2 scheduler + qmd serve integration. Once one of those lands this becomes trivial to resolve.

[jaylfc]

Unblocked by the framework-agnostic runtime rework (commit 9e2759e, design doc docs/design/framework-agnostic-runtime.md). The last concrete piece needed before this issue can move is #55 — a single host-side memory service that the scheduler can route to. Once that lands, this becomes 'call the scheduler with memory target instead of calling QMD directly'.