Add ACCELERATED_COMPUTE health check for the MPS control daemon

[harishxr]

Summary

Add an ACCELERATED_COMPUTE doctor health check that reports the instance IMPAIRED when the NVIDIA MPS control daemon is persistently unavailable. Today the per-task MPS health gate refuses individual MPS tasks when the daemon is down, but nothing tells the control plane the instance itself is unhealthy, so it keeps routing MPS tasks to a box that rejects them one by one. This check surfaces a persistently broken daemon as an instance-level health signal on DescribeContainerInstances and the "ECS Container Instance Health Change" EventBridge event.

Implementation details

• agent/doctor/mps_daemon_healthcheck.go - the health check. Runs the MPS control-daemon probe once per doctor tick and carries its own consecutive-failure counter: it reports IMPAIRED only after 3 consecutive failures and resets on a single success, so a brief RestartSec=1 systemd bounce does not flap. GetHealthcheckType() returns ACCELERATED_COMPUTE. Reuses the existing MPS probe (pipe-directory pre-check + one ProbeControlDaemon, 3s timeout); a nonzero exit or a timeout (wedged daemon) both count as a failure.
• agent/app/agent_mps_healthcheck_unix.go - registers the check on MPS-capable instances via the shared gpu.ShouldAdvertiseMpsCapability predicate.
• agent/app/agent_mps_healthcheck_unsupported.go - no-op so non-Linux builds are unaffected.
• agent/app/agent.go - wiring in newDoctorWithHealthchecks.
• agent/app/agent_capability_unix.go - extracted mpsCapabilityInputs() so the gpu-sharing-mps capability and the health check decide from the same host facts (single source of truth). No behavior change to the capability.

Testing

New tests cover the changes: yes

Live E2E on a EC2 GPU instance:

• systemctl stop nvidia-mps.service: failures 1 and 2 stay Ok, failure 3 → IMPAIRED, after restart → OK.

Fault → IMPAIRED was ~75s (3 failures at the ~25s heartbeat tick); recovery is immediate on the first successful probe.

DescribeContainerInstances --include CONTAINER_INSTANCE_HEALTH 

"healthStatus": { "overallStatus": "IMPAIRED", "details": [
    { "type": "ACCELERATED_COMPUTE", "status": "IMPAIRED", "lastStatusChange": "..." },
    { "type": "CONTAINER_RUNTIME",   "status": "OK" } ] }

The "ECS Container Instance Health Change" EventBridge event 

(capacityProviderName: null, ec2InstanceId set, full healthChecks[]):
{ "detail-type": "ECS Container Instance Health Change", "source": "aws.ecs",
  "detail": { "capacityProviderName": null, "ec2InstanceId": "i-...",
    "overallStatus": "IMPAIRED",
    "healthChecks": [ { "type": "ACCELERATED_COMPUTE", "status": "IMPAIRED", ... },
                      { "type": "CONTAINER_RUNTIME", "status": "OK", ... } ] } }

Description for the changelog

Enhancement - Add ACCELERATED_COMPUTE health check for the MPS control daemon

Additional Information

Does this PR include breaking model changes? If so, Have you added transformation functions? No

Does this PR include the addition of new environment variables in the README? No

Licensing

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[amogh09]

Opportunity to use table driven tests?

[harishxr]

Added in fe5d32c