feat: add X-ASR + Qwen3-ASR backends, Gemma-4 LLMs; fix review findings

add gemma 270m for test use

Fix llama-cpp-python blocking on HF Spaces due to excessive threads

Optimize speaker diarization performance for long audio files

default asr backend to sensevoice 2024

bugfix to permission err on model

add sensevoice small as 2nd ASR backend

Update src/utils.py

Update src/utils.py

Create utils.py