Context Agent poller resilience PR

[dberardo-com]

DISCLAIMER: this is PR is vibe coded !!

Relates to #266

Summary

• Keep and cancel the poller as part of the FastAPI application lifespan.
• Restart its coroutine after unexpected termination (5s backoff).
• Bound next_task() with CONTEXT_AGENT_CLAIM_TIMEOUT_SEC (default 60s).
• Wake immediately on /trigger, with configurable idle fallback
  CONTEXT_AGENT_IDLE_POLL_SEC (default 30s).
• Isolate unexpected task-handler exceptions so they do not terminate the poller.
• Log local claim errors, unexpected handler failures, and task durations,
  without logging user prompts or task payloads.
• Do not log empty 2-second polls.

Scope

Only ex_app/lib/main.py is needed for this source change.
Do not include site-specific ConfigMap patches (agent.py, context_chat.py,
nc_model.py), credentials, or Kubernetes cleanup logic in the upstream PR.

Verification

• python -m py_compile ex_app/lib/main.py
• An isolated asyncio smoke test passed: trigger wake, handler-exception
  containment, supervisor restart, claim timeout/recovery, cancellation.
• Full integration tests against the upstream dependency environment are still
  required; the isolated test is not equivalent to a full application test.
• User's local patched deployment demonstrated LocalAI failure followed by
  successful processing of a new task without pod restart. This does not prove
  unpatched upstream would fail the same test.

Limitations

• A failed LLM inference is not automatically retried.
• asyncio.wait_for depends on cooperative cancellation and can leave a
  server-side claimed task in RUNNING if the response is lost or arrives late.
• Pod crashes can still leave TaskProcessing RUNNING records; no reconciliation
  is provided here.
• Unexpected poller failure while a TaskGroup is active may cancel its child
  handlers; this is safer than leaving the poller permanently dead but does
  not guarantee recovery of in-progress tasks.
• An active task that gets stuck inside handle_task() is not forcibly timed
  out by the task-acquisition timeout.
• The existing 2-second busy polling is retained but logged only on errors;
  it still generates Nextcloud API traffic.