Replace resampy with pysox for audio resampling

Nobody has claimed this yet.

Turns out resampy is SLOW. We could speed openl3 quite a bit by swapping it out with pysox, resample the audio into a tempfile and load the templfile for further processing.

```python

with tempfile.TemporaryDirectory() as tempdir:
filename48 = os.path.join(tempdir, os.path.basename(os.path.splitext(file_path)[0]) + "_48.wav") 
# resample to 48000 Hz
tfm = sox.Transformer()
tfm.convert(samplerate= 48000, n_channels=1)
tfm.build(file_path, filename48)

# load the audio file
sr = 48000
audio, srload = librosa.load(filename48, sr=None)
assert srload == sr
```

NOTE: this example assumes wav input, and would have to be adapted to support various formats (in particular mp3)

Start by locating where resampy is used in the audio-loading and resampling path, then review the existing dependency and input-format handling. Replace the resampling path with pysox while preserving support beyond WAV, especially MP3, and verify that the resulting audio is loaded at 48000 Hz with one channel.

Written by the indexing model from the issue text.