[LauraGPT]

Summary

Keep #3797's six empty-token batch regressions active in both NumPy CI matrices. Add the core Paraformer model and regression file to both pull-request and main-push filters. No inference code changes here.

Validation

• Before this workflow change: actual pytest collection selects 199 tests and 0/6 empty-batch regressions; the selection contract fails.
• After: collection selects 205 tests, including all six empty-batch cases; both event filters cover the model/test paths.
• The identical production/test source was already validated for fix(paraformer): keep one keyed result per utterance when no tokens are predicted #3797: 6 CPU failures before the fix, 205 focused tests passing after it; 32 real L20Z SANM/CIF synthetic cases change from invalid to correct results. FP32 and BF16-autocast encoder kernels were observed; CIF retains its own FP32 path.
• These are scoped tests, not pretrained ASR quality, full repository tests, VAD/speaker model inference, or Nano leaderboard work. No new model weights were downloaded. The broad suite's unapproved model/data integrations were not run.
• Exact new-head CI and its two downloaded NumPy artifacts must contain all six regressions before normal merge. No new package release is implied.