feat(audio): support explicit PCM16 input without container probing

[LauraGPT]

Summary

Related to #3739; this does not claim to reproduce or diagnose the reported three-minute decoder stall.

• Add opt-in input_format="pcm_s16le" for known mono signed little-endian PCM16 bytes. It normalizes samples directly, bypassing container detection/decoding, and rejects partial two-byte samples. The default auto byte-format behavior is unchanged.
• Forward the option through data preparation, explicit byte batches, multimodal inputs, ordinary inference, VAD preprocessing and export input preparation. Include matching English/Chinese API examples.
• Fix a pre-existing sample-rate propagation problem exposed by this path: VAD receives the source rate; ASR receives the rate of the already-resampled segments. Keep the original input rate stable across a batch so a prior ASR call cannot alter how subsequent input is loaded.
• Run the new regression suite and existing audio-byte tests in both existing NumPy CI lanes; add their source/test trigger paths.

result = model.generate(input=pcm_bytes, input_format="pcm_s16le", fs=16000)

The format flag does not make other encodings valid PCM16. Do not use it for encoded WAV/MP3 bytes, floating-point PCM, other bit depths, big-endian audio or interleaved stereo. Streaming callers retain their cache/chunk/final-chunk handling.

Validation

• Initial format regressions on unchanged main: 17 failed / 2 existing-behavior controls passed; implementation then passed the same tests.
• Additional 8/48 kHz VAD tests covering arrays/PCM, single/batch, and constructor/per-call rates: 16 failed before the rate fix, then passed. They exercise actual preprocessing/resampling with recording model doubles, not downloaded acoustic models.
• Final focused input, byte decoder, API documentation/signature and submodel tests: 93 passed, 0 skipped.
• Full existing NumPy workflow test selection plus the added suites: 93 passed, 0 skipped per environment, on NumPy 1.26.4 and 2.4.0, each with PyTorch 2.10.0+cpu.
• Independent final review found no remaining P1/P2 issues; git diff --check passes.

Coverage includes probe/decoder bypass, signed sample boundaries, malformed input, default behavior, batch/multimodal routing, nonempty VAD segments, streaming options, per-call reset, export preparation, and bilingual executable examples. Export preparation uses a capture boundary, not an ONNX export acceptance test. No model weights, GPU, model-quality evaluation, original reporter-environment reproduction, package release or website deployment is claimed.

Integration Status

Merged as bb303f85ea415e64d2188de6847ddcf2adc2593f; merge parents and tree were verified against the tested candidate 2d6cb0ed216254b2f7d046fb78027a3933e3ddc2.

• All four exact-head hosted workflows passed: NumPy compatibility, KWS, MOSS, and product site.
• The CI follow-up explicitly installs/checks FFmpeg. Downloaded JUnit artifacts from NumPy run 36888888302 confirm 93 tests, 0 failures/errors/skips in each lane. The older head had a missing-FFmpeg skip and is not used as zero-skip evidence.
• Source-only change; not a new PyPI release. Issue Bug: model.generate(input=bytes) hangs for minutes on raw PCM chunks that trigger audio container false-positive #3739 remains open because its original decoder stall has not been reproduced.
• All five exact-merge workflows passed: NumPy compatibility (36893811611), KWS (36893811800), MOSS (36893811789), product site including browser checks (36893811808), and API documentation (36893811759). Downloaded merge JUnit artifacts again confirm 93 tests and 0 failures/errors/skips per NumPy lane.