Centralize handling AI Assistant messages failures in domain layer

Description

Currently, message failure handling for AI Assistant chat is split between the worker (MessageProcessor) and Oban telemetry handlers. This means:

• Some failures (Apollo HTTP timeout, validation errors) are handled inside the worker.
• Others (process crash, Oban job timeout/discard) are handled via Oban telemetry events.

This works today, but it:

• Couples core domain logic (marking messages failed) to Oban-specific events.
• Scatters the failure-handling code across two modules.
• Makes it harder to swap Oban for another execution mechanism in the future.

Proposal

• Extract a domain-level helper (e.g. Lightning.AiAssistant.Messages.transition/2) that:
  • Updates message status (:processing, :success, :error, etc.).
  • Stores failure cause (timeout, HTTP error, etc.).
  • Broadcasts status changes.
• Call this helper from both the worker and telemetry handlers.
• Make telemetry a “catch-all” for failures not caught inside the worker.
• Optionally store failure_category and failure_message in the DB for observability.

Benefits

• Single source of truth for message state transitions.
• Decouples domain logic from Oban.
• Easier to test.
• Easier to replace Oban in the future without touching core message logic.

Start by reading the MessageProcessor worker and the Oban telemetry handlers to map their existing failure paths and status updates. Define the domain-level transition helper and confirm how both callers would use it; done means message transitions, failure causes, and broadcasts have one shared path, with any persistence and tests agreed.