fix(runtime): skip empty deduplicated event segments

[LauraGPT]

Summary

Fix an existing IndexError in the ONNX and libtorch SenseVoice rich-transcription postprocessors when consecutive language segments contain the same event and a segment has no spoken text.

After duplicate leading-event removal, a segment can become empty; both runtime copies then index it in get_event(). Add the same post-deduplication empty-segment guard already present in core FunASR. Keep the current event/emoji state and subsequent transcript intact.

• Two-line guard in each runtime copy; no dependency, model, inference or event-mapping change.
• Parameterized tests load all three actual postprocessor modules and cover repeated event-only segments, following text, event-only input without speech, and distinct events.
• Independent of fix(sensevoice): stop rendering the Cough event as the Sneeze emoji #3747's Cough mapping fix. Both fix: keep token separators in timestamp_sentence raw_text #3746 and fix(sensevoice): stop rendering the Cough event as the Sneeze emoji #3747 merged during this work; the branch includes actual main 217d10aa99587dc9d7b42b59641a856716661cbc and preserves both changes.

Validation

• Before the fix, the new 30-case regression file gives 18 failed / 12 passed: only ONNX/libtorch duplicate-event cases raise IndexError; the core control and distinct-event controls pass.
• The same regression file against exact fix(sensevoice): stop rendering the Cough event as the Sneeze emoji #3747 head also gives 18 failed / 12 passed, demonstrating this bug predates and is not caused by its mapping change.
• After the guard: 30 regression cases passed; combined with related speaker/emotion documentation checks, 61 passed.
• After merging actual current main, native CPU run below: 71 passed, zero skips/errors, 11.95 seconds test time (13.24 seconds enclosing command).

python -m pytest -q -p no:cacheprovider \
  tests/test_rich_transcription_event_segments.py \
  tests/test_rich_transcription_postprocess.py \
  tests/test_timestamp_tools.py \
  tests/test_speaker_emotion_docs.py

The timestamp tests import the real FunASR package; the postprocessing tests use the repository's file-loader pattern to execute each shipped module without initializing a runtime engine. This is postprocessor/native helper coverage, not ONNX/libtorch model inference, a model download, a full test-suite pass or a package release. CUDA was disabled and model access offline.

Changed Python files compile; effective diff passes whitespace checks. Independent read-only review of the final current-main diff found no P1/P2. Signed commits include DCO sign-offs. No hosted CI result is claimed at submission.

Integration Status

Merged as 800f7565e8e850b079b9e111d12107c19fc6121f. The actual merge tree exactly matches the reviewed and locally tested signed head d7a33c862c9d03a68ff416110416921e49728319; parents and tree were checked after merging.

The exact-head ONNX package workflow 36817570899 and exact-merge workflow 36817826794 both completed successfully for Python 3.11 and 3.12. These are package build/install/wrapper checks; the 71 focused postprocessing/timestamp/documentation tests above are separate native evidence, not a claim that the hosted workflow runs that full selection.

This change is now in repository source. No Python package/model release, website deployment or running inference-service restart was performed.