Cactus VAD is 15% slower than ONNX runtime

- Tested in `M4 Pro`.
- `cargo bench -p vad --all-features` for running https://github.com/fastrepl/char/blob/main/crates/vad/benches/vad.rs
- `silero-rs` use `ONNX runtime`(https://github.com/microsoft/onnxruntime, through https://github.com/pykeio/ort)
- `silero-rs` use [1.8 MB model](https://github.com/emotechlab/silero-rs/blob/main/models/silero_vad.onnx)
- Unzipped https://huggingface.co/Cactus-Compute/silero-vad/blob/main/weights/silero-vad-int8.zip is `664 KB`

Start with crates/vad/benches/vad.rs and run `cargo bench -p vad --all-features` as described. Compare the Cactus VAD benchmark with silero-rs using ONNX Runtime, accounting for the different model sizes; done means identifying and addressing the source of the reported performance gap and verifying it with the benchmark.

Written by the indexing model from the issue text.