v6.2.3

torchaudio is now an optional dependency instead of required. Only torch is required — the model itself never used torchaudio, only the `read_audio` / `save_audio` helpers did.

Users who need audio I/O must now install a backend explicitly:

- `pip install silero-vad[audio]` (torchaudio)
- `pip install silero-vad[codec]` (torchcodec)
- `pip install silero-vad[all]` (everything)

Key changes:

- `read_audio` / `save_audio` now work with either torchaudio or torchcodec; with neither installed they raise an error naming the extra to install
- Fixed: `import silero_vad` failed when onnxruntime was not installed — the sequence path is now imported lazily
- numpy is now declared in [onnx-cpu] / [onnx-gpu]

If you load audio yourself, no backend is needed at all:

```python3
from silero_vad import load_silero_vad, get_speech_timestamps

model = load_silero_vad()
audio = ...                                      # 1-D float32 torch.Tensor from any source
ts = get_speech_timestamps(audio, model, return_seconds=True)
```