Add MPS control-daemon health probe

[harishxr]

Summary

Adds a functional health probe for the NVIDIA MPS control daemon to the shared lib. ProbeControlDaemon execs the MPS control utility with a control command and reports whether the daemon is serving, distinguishing three outcomes: serving (exit 0), not serving (positive exit code from the utility), and wedged (the utility hangs and is caught by a timeout).

This is the first PR in a stack that adds an MPS daemon health gate to the agent:

1. This PR: the probe primitive.
2. An MPSDaemonResource TaskResource that wraps the probe in a bounded retry and gates task launch.
3. Wiring that attaches the resource to MPS tasks only.

ProbeControlDaemon is exported but has no caller in this PR by design.

Implementation details

• ProbeControlDaemon(exec, command) runs /usr/bin/nvidia-cuda-mps-control under a 3s context timeout, feeding the control command on stdin, and returns a ProbeResult{ExitCode, Stdout, Latency, TimedOut, Err}.
• A wedged daemon accepts the socket connection but never replies, so the utility hangs. The context timeout converts that hang into TimedOut=true, which is the failure mode systemd liveness cannot detect. Exit code, output, and latency are captured so a caller can surface them as a task stop reason.

Testing

• The probe was tested end to end as part of the health-gate integration testing: a healthy daemon passed, a stopped daemon was refused via the exit-1 path, and a SIGSTOP-wedged daemon was refused via the timeout path (probe returned TimedOut=true at ~3s per attempt while systemd still reported the unit active).
  New tests cover the changes: yes

Description for the changelog

Enhancement - Add MPS control-daemon health probe

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.