fix(vad): exclude idle time from dynamic silence schedule

[LauraGPT]

Summary

Keep the dynamic FSMN silence schedule tied to the current detected speech segment instead of time spent waiting for speech. Related to #3302; this PR deliberately does not close that report.

• Track total input samples and derive elapsed speech time from the VAD's absolute start timestamp. Initial and inter-utterance silence no longer ages the next utterance; fractional-millisecond packets retain their samples.
• Preserve backdated starts and the last open segment when one feed returns several signals.
• Reset stream state on the next feed after EOS, while preserving final state for the realtime service's tail fallback. Repeated finalize() is idempotent.
• Leave the silence schedule, noise threshold, compensation, model weights and ASR decoding unchanged.

Verification

On ind-gpu8, using PyTorch 2.8.0+cu128 with CPU FSMN inference:

python -m pytest -q tests/test_dynamic_streaming_vad.py \
  tests/test_realtime_ws_service.py tests/test_realtime_ws_benchmark.py \
  tests/test_fsmn_vad_dynamic_silence.py tests/test_fsmn_vad_streaming_buffers.py

122 passed, no skips. The wrapper suite includes 16 tests. The original idle-duration defect and both EOS compatibility defects were reproduced as failing assertions before their respective fixes. git diff --check passes; independent static review found no remaining actionable issue.

Real-model controlled check: pinned FSMN snapshot v2.0.4, first 30 seconds of the reporter's YT-02.mp3, mono 16 kHz, 256 ms transport packets, three seconds of trailing silence. Compare the same clip with zero vs 60 seconds of preceding silence:

Wrapper	No initial silence	60 seconds of initial silence
Before	2 segments	3 segments
After	2 segments	2 segments

After subtracting the prefixed silence, the fixed segment boundaries are [[0, 17690], [24140, 30280]] and [[-220, 17690], [24140, 30280]] ms. The 220 ms onset difference is VAD lookback into the prefixed silence; not every boundary is asserted identical.

Scope And Follow-Up

This uses the original playback file, not the reporter's captured microphone PCM, and does not measure ASR transcription accuracy. It establishes an idle-time-dependent VAD segmentation defect, not complete resolution of microphone truncation. The 42-second decode warning in the report remains a separate investigation. No arbitrary transport-packetization invariance or hardware fix is claimed; #3302 stays open for real microphone validation.