speech-adjacent

Introduces AnySimLite, a lightweight similarity encoder for on-device speech-adjacent classification tasks, achieving state-of-the-art or competitive performance while using less than 1/250th the model size of the qLLaMA-LoRA-7B baseline.