fix(kws): preserve streaming frames across packet and EOS boundaries

[LauraGPT]

Summary

• Preserve all valid KWS frames across arbitrary packet boundaries, including short input and empty EOS. Nonfinal public calls return [], without marking encoder tail state prematurely.
• Use canonical frontend and encoder windows; preprocess initial real context through normalization/scaling/positions exactly once. Flush only uncommitted frames at EOS.
• Avoid sending empty input through real SANM convolutions when zero right context remains.
• Fix shared WavFrontendOnline EOS for zero valid fbank frames and short final LFR context.
• Preserve nonfinal internal generate_chunk None and optional output files; clarify the public EOS-only result contract.

Verification

• Before fix: the initial seven-method regression reported 10 failing subcases and 25 errors; an additional real SANM convolution test reproduced the zero-context EOS crash.
• Final and exact signed-head CPU runs: 40 tests, 184 subtests passed across KWS plus related frontend/VAD tests.
• Tests retain actual AutoModel configuration resets, audio frontend, positional encoding, SCAMA overlap and a small real SANM convolution layer. Deterministic context-sensitive windows verify partition consistency.
• Official iic/speech_sanm_kws_phone-xiaoyun-commands-online weights: all checkpoint keys matched. On the same positive WAV, candidate file/array/960-sample/irregular packets have identical 76x256 encoder-frame hashes and keyword result. Synthetic silence is rejected. The baseline complete-file positive also detects the keyword; scores change, so no accuracy improvement is claimed.
• Independent source/log review completed; runtime source hashes, logs, audio/weight SHA256s and rollback bundles are retained. Source commit is signed with DCO.
• Expanded Python 3.11 CPU CI includes all six regression files.

Boundaries

Results remain utterance-final, not per-packet wake events. Accumulated encoded output still grows with utterance length, so callers must send EOS. Direct generate_chunk final calls intentionally include remaining right context. Export wiring is unchanged but export execution was not run. Actual-model evidence is a CPU functional smoke comparison of one positive sample plus synthetic silence, not a natural-negative, accuracy, latency, microphone or GPU benchmark. No PyPI release, website deploy or issue closure is part of this PR.