fix(audio): preserve ffmpeg fallback sample rate

[hulkbig]

Summary

The FFmpeg fallback already returns audio at fs, but load_audio_text_image_video retains the source audio_fs and resamples the decoded waveform again. Forcing the fallback on a generated one-second, 48 kHz WAV with fs=16000, audio_fs=48000 produces 5,334 samples on the base commit instead of 16,000. This change returns the expected 16,000 samples.

• Record audio_fs = fs after a successful FFmpeg fallback.
• Pass a separate input_sr for raw .pcm files so their source rate is preserved; direct helper calls still default input_sr to sr. Container source rates still come from their headers.
• Add regression coverage for container paths/bytes/BytesIO, consumed-stream rewind, decoder controls, and raw PCM rates. Simply adding audio_fs = fs without the PCM input-rate change fails five of the six raw-PCM controls.

Type of change

• Bug fix
• Documentation
• Example or demo
• Runtime or deployment
• Benchmark or evaluation
• Model/training change

Validation

• python -m compileall funasr examples tests
• Docs or links checked (not applicable)
• Runtime/deployment command tested (not applicable)

The existing CPU CI test selection was run locally with NumPy 1.26.4 and 2.4.0. Each run: 238 passed, 6 skipped, 10 subtests passed. The six skips are real torchaudio decoder controls requiring optional torchcodec, which the CI environment does not install.

OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMBA_NUM_THREADS=1 HF_HUB_OFFLINE=1 python -m pytest -q \
  tests/test_numpy_compatibility.py tests/test_numpy_cluster_compatibility.py \
  tests/test_eend_feature_compatibility.py tests/test_pypi_metadata.py \
  tests/test_sensevoice_word_timestamps.py tests/test_pcm_input_format.py \
  tests/test_load_audio_bytes.py tests/test_timestamp_tools.py \
  tests/test_punc_model_none.py tests/test_paraformer_hotword_file.py tests/test_cli.py

With optional torchcodec==0.10.0 installed, python -m pytest -q tests/test_load_audio_bytes.py tests/test_pcm_input_format.py -k 'ffmpeg or normal_decoders' passed all 64 tests, with no skips, including real first-decoder controls. The complete two-file suite in that environment returned 146 passed, 1 failed: TestLoadAudioBytes::test_decodes_big_endian_rifx_wav. That endian-decoding failure reproduces identically on untouched base e7e61293d308f4342ab9307895e710e4b51b949c.

pip check, the changed audio-byte test file's Black check, and git diff --check passed.

User impact

Preserves audio duration and sample rate when normal decoders fall back to FFmpeg, while retaining raw-PCM source-rate behavior.

Notes for reviewers

This is the earlier audio-loader fallback boundary, separate from the speaker-boundary resampling addressed in #3763. No GPU, model weights, or acoustic-quality evaluation was used. Hosted CI also passed both NumPy matrix jobs on 1d5fd87d933118198075f358dddc7644bc063b42: NumPy 1.26.4 and NumPy 2.4.0.