Add Relay primitives for actors, boundaries, and capability grants

[zaridan]

Represent Relay as core conversation infrastructure, using the actor and event foundations already in the repository.

Current foundation:

• Human Agents already exist.
• Assistant actors and assistant tokens already exist.
• Threads already record author_kind and actor identity.
• Draft-before-send already proves policy-gated AI action.
• Webhooks and event outbox already provide extension delivery.

Remaining work:

• Model Relay requests without duplicating existing Agent or Assistant concepts.
• Model boundaries between internal operators, external participants, systems, vendors, and future channels.
• Model capability grants: who is allowed to ask whom to do what, for which conversation, under which policy.
• Keep Relay channel-neutral. Email may be the first proof path, but Relay must not depend on Gmail, IMAP, SMTP, or email headers.

Acceptance:

• Storage/types exist for Relay requests, participants/boundaries, capability grants, and actor references.
• Relay records explicit authority and provenance for every state transition.
• The model supports human, assistant, system, and message-originated flows.
• No first-party or commercial module receives private core capabilities.
• No hosted Relay broker or Resonant IQ data path is introduced.