tests: accept staged Cray MPICH in Comm_Backends provenance

[dannybaths]

Summary

This PR adds an MPI PROVENANCE: verdict to Pytorch_Comm_Backends_Check and
uses that verdict as the CTest pass criterion.

The existing GPU-AWARE MPI: line is left unchanged as the strict OpenMPI
diagnostic. The new provenance verdict applies Bob's 3 September rule: pass
when PyTorch is using either GPU-aware OpenMPI from the loaded module or a
staged Cray/MPICH-family MPI, and fail when it resolves to OS/system MPI or an
unrecognised MPI.

Why

After PR187, Tcl Environment Modules can be parsed correctly, but AAC7 still
fails the original gate honestly because the site PyTorch is linked to staged
Cray MPICH rather than the loaded OpenMPI module. That is not an OS MPI leak.

The new gate keeps the OpenMPI diagnosis visible while changing the CTest
decision to the intended provenance rule.

Validation

• AAC7 job 10537: standalone Pytorch_Comm_Backends_Check passed with
  MPI PROVENANCE: OK for staged Cray MPICH.
• AAC7 job 10539: focused Phase C in-harness run passed with the new CTest
  gate, Regex=[MPI PROVENANCE: OK].
• AAC6 job 19659: Lmod/OpenMPI no-regression proof passed. The stock output
  before the new provenance line was identical; with this PR's CTest rule, the
  added MPI PROVENANCE: OK line matched the pass expression and the test
  passed.

[bobrobey]

Good to get this to pass with Cray MPICH