ci: run rich transcription regressions in ONNX package checks

[LauraGPT]

Summary

• Run the 36 existing rich-transcription regressions for core, ONNX and libtorch in the existing ONNX package workflow, on both Python 3.11 and 3.12.
• Trigger that workflow for edits to the core/libtorch postprocessors and either regression test file, in addition to the existing ONNX paths.
• Install pytest as CI tooling only. Preserve all wheel/sdist checks, the no-Torch install assertion, the matrix and read-only permissions.

This closes the CI coverage gap after #3747 and #3748: those regressions were verified natively but were not executed by the hosted package workflow.

Validation

• Structured workflow audit fails on the unchanged workflow for the missing paths, then passes with this change.
• The exact new pytest command: 36 passed, with no Torch installed and no models downloaded or inference performed.
• Existing release contract: 5 passed.
• git diff --check passes.

Native validation used Python 3.12. Hosted PR validation passed on Python 3.11 and 3.12; each job log explicitly reports 36 passed, and all existing package checks passed.

Merge Status

Merged as 4a84879ae8ca4d2430953b11c49d609217d93988. The merge tree matches the validated head. Main-branch validation also passed on both Python versions.

This changes CI only; it does not publish a package or deploy a website.