ci: cover CLI regressions in CPU NumPy checks

[LauraGPT]

Summary

Run the existing CLI regressions in both CPU NumPy compatibility lanes, and trigger that workflow when either funasr/cli.py or tests/test_cli.py changes.

The gap became visible while reviewing #3750: the CLI tests exist and can catch user-facing output regressions, but no current workflow explicitly selects them or includes their source/test paths. This change reuses the existing Python 3.12, CPU PyTorch 2.10.0, NumPy 1.26.4/2.4.0 environments and artifact reporting.

• Add both CLI paths to pull-request and main-push filters.
• Execute tests/test_cli.py alongside all existing suites, and name CLI coverage in the step label.
• Preserve dependencies, permissions, matrix, timeouts, FFmpeg setup, failure propagation, and uploaded evidence.

Scope: one existing workflow only. No CLI implementation or test changes; this does not include or resolve #3750, change branch protection, or publish a release.

Validation

• Structured YAML audit fails on the original workflow for the four missing trigger paths and missing pytest selection, then passes on this change.
• All workflow shell blocks pass bash -n; git diff --check passes.
• Independent source review found no P1/P2 issues.
• Native NumPy 1.26.4 and 2.4.0 combined suites each passed: 120 JUnit cases, including 27 CLI cases, with 0 failures/errors/skips. Pytest reports 118 passed plus 2 passing subtests; three existing dependency/frontend warnings remain.
• Exact-head hosted run 36931513638 and exact-merge run 36931894062 both passed. Downloaded JUnit artifacts for all four lanes confirm 120 cases, including 27 CLI cases, and 0 failures/errors/skips per lane.

Merged as 2ddbd3d9970af621f8bad796a3ab12939218a795. Merge parents and tree were verified against the signed tested head cfc8c47708cf093438bf097a5eab86ac90fcd60a; no runtime code or package release changed.

CLI tests exercise the real command module with temporary files and existing model doubles. No model downloads, acoustic inference, GPU execution, or end-to-end recognition-quality claim.