fix: set context.task for scheduler-run tasks so end-of-turn extensions fire

[King0James0]

Fixes #1776.

The scheduler runs tasks by calling agent.monologue() directly, so context.task is never assigned for scheduler-run contexts — and agent.py gates both message_loop_end and monologue_end extension dispatch on context.task and context.task.is_alive(). Net effect: no end-of-turn extensions (including the built-in _memory memorize extensions) ever fire for a Scheduled/AdHoc/Planned task.

The scheduler already wraps each run in its own DeferredTask (registered in _running_deferred_tasks before the wrapper starts); this assigns it to context.task right after the run's context is resolved — the same contract the normal chat path establishes via run_task(). 4-line diff in _run_task_wrapper.

Effects:

• message_loop_end / monologue_end now dispatch for scheduler runs, so memorization and extension-based plugins see scheduled work. A killed run still correctly skips them — killing the task makes is_alive() false, so the post-mortem guard's intent is preserved.
• context.is_running() is now truthful for scheduler-run contexts.
• A user message sent to a task context mid-run routes as an intervention (like a normal chat) instead of spawning a second concurrent monologue on the same context.

The if not (context.task and context.task.is_alive()) guard leaves the context untouched in the edge case where a live task already exists on it (e.g. a user actively chatting in the task context when the run starts).

Verified on a v2.4 container: stock = a probe extension on monologue_end never fires for an ad-hoc scheduler run (while a normal chat turn fires it); patched = the same run fires message_loop_end + monologue_end with the task's context id within seconds, and a killed run fires neither.