[LauraGPT]

Summary

Streaming Paraformer currently reads feats.shape[1] even when the real
online frontend returns a one-dimensional empty array before its first LFR
frame. This raises IndexError for short initial packets and no-feature calls.
This was explicitly left outside the VAD fix in #3822.

• Check feature presence with feats.size without changing the nonempty path.
• Keep frontend audio accumulation/reset and the existing cached-final flush.
• Add seven regression methods to the already-run standalone ONNX test file.

Validation

• Current-main red reproduction: 5 failing / 7 passing test methods;
  all five failures were IndexError: tuple index out of range.
• Candidate wheel built from sdist, twine metadata and private install verified;
  all 71 related package/rich-transcription tests pass, no skips.
• Real NVIDIA L20Z / ONNX Runtime GPU1.23.2, NumPy1.26.4, kaldi-native-fbank1.22.3:
  32 new boundary scenarios, baseline25IndexErrors/7passes to candidate32passes.
  The seven existing successful scenarios retain identical outputs and cache
  fingerprints. Candidate encoder/decoder profiles contain162CUDA node events
  and noCPU node events.
• FP32,batch1,16kHz,chunks[5,10,5], deterministic silence/sine inputs; reused
  hash-verified synthetic ONNX encoder/decoder graphs, actual frontend/CIF/cache/
  decode/result path. This is runtime contract validation, not pretrained ASR
  accuracy, full repository-suite coverage, BF16, or leaderboard validation.
• Coverage includes first packets0/1/160/320/399/400/480/560/879samples, repeated
  160-sample accumulation, empty final after buffered features, existing-cache
  final/nonfinal empty calls, stream reuse, and880/960/9600-sample controls.

Only two files change. No model downloads, shared environment changes, package
version bump, release, or PyPI upload. Published0.4.4 does not yet contain this
new repair; this PR is a source fix for the next standalone package.