feat: make load shedding queue- and priority-aware

[theo372001]

Summary

Closes #777.

Load shedding previously used a single flat Celery queue-depth threshold
and only checked LLM provider health for the synchronous verify route,
relying on coarse request-level thresholds alone. This tunes it with
queue depth, provider health, and job priority signals:

• Priority-scaled queue threshold - services/load_shedder.py now
  scales the queue-depth threshold by the job's declared priority
  (low=0.5x, normal=1x, high=1.5x, urgent=3x), so low-value
  backlog is shed well before high-priority work. The middleware
  pre-check (evaluate_load_shed) stays a cheap, priority-agnostic
  outer gate since it runs before the request body is parsed; the
  precise, priority-aware check runs in ensure_queue_capacity,
  wired up from tasks.create_task right before a job is queued.
• Provider health for async job creation - previously provider
  health was only checked on the synchronous /ai/humanitarian/verify
  route. It's now also checked when queuing async jobs via
  /v1/ai/inference with type=humanitarian_verification, so
  LLM-dependent work is shed up front instead of being queued, retried,
  and eventually dead-lettered.
• Metrics - both paths reuse the existing LoadShedError ->
  ErrorEnvelope plumbing (so shed responses stay consistent with the
  rest of the service's error envelope), and queue-pressure sheds are
  now additionally broken out by priority tier in a new
  queue_shed_by_priority_total{priority} counter alongside the
  existing requests_shed_total{reason,method,endpoint}.
• Documented the tuned behavior in app/ai-service/README.md.

Test plan

• tests/test_load_shedder.py - 22 new tests (30 total): priority
  multiplier resolution, priority-scaled queue-depth thresholds
  (low sheds sooner, urgent tolerates more), provider-aware job
  creation (sheds LLM-dependent types, ignores non-LLM types,
  queue check short-circuits before the provider check), the new
  priority-labeled metric, and two end-to-end HTTP tests proving
  envelope consistency for both new shed paths.
• Full pytest suite run locally - no regressions in any test
  touching load shedding, tasks, or metrics.
• flake8 --select=E9,F63,F7,F82 clean on all changed files.

🌊 Stellar Wave application: 668468d4-8d29-4f21-a1a1-9d2b5736ddbb

[drips-wave]

@theo372001 Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits

[Cedarich]

@theo372001 please resolve conflicts

[Cedarich]

@theo372001