process function should be deterministic.

[WenqingZong]

Currently we have some non-deterministic behavior in process function:

        for i in 0..num_chunks {
            let sample_range = if i < num_chunks - 1 {
                (self.processed_samples - self.deleted_samples)
                    ..(self.processed_samples + vad_segment_length - self.deleted_samples)
            } else {
                (self.processed_samples - self.deleted_samples)..self.session_audio.len()
            };
        // The rest of process function...
        }

The remainder size changes a lot, and this leads to an unsuccessful PR that checks the returned current_speech indeed include unprocessed samples - but now unprocessed is always 0.

A desired modification is deleting the last for iteration. If there are unprocessed samples, just keep them in internal buffer and wait for more samples to come.