[benoitc]

Lets a task outlive the node, for hornbeam's durable executions.

• barrel_mcp_task_store behaviour (task_store app env, ETS default). With a durable backend, a task left working on a restarted node becomes failed ("Task interrupted by a restart"); a modern input_required task resumes through tasks/update.
• barrel_mcp_task_provider behaviour: the application owns id, state, input rounds and cancel. A tool registered with task_provider => Module returns {task, TaskId}. tasks/* and subscriptions/listen look in the store first, then ask each provider, so nothing about the id is kept by barrel_mcp.
• barrel_mcp_tasks:principal/1 (restart-stable id for the owner's principal) and changed/3 (status notifications for a hosted task).
• barrel_mcp:task_owner/1, task_allowed/1, register_task_provider/1.
• Fix: terminal tasks with a ttl were never evicted by the sweep.

Default store and no provider: behaviour unchanged. Guide: guides/durable-tasks.md.