Deploy asr-whisper-large-v3-salt on Modal

[huwenjie333]

This PR adds the script to deploy the asr-whisper-large-v3-salt on Modal platform.

• The vLLM inference engine was attempted and it doesn't improve the performance, so we revert back to the transformer pipeline methods. See the attached screenshot.
• currently dynamic batching feature is disabled by setting the batch_size=1, so that any request will be processed immediately without waiting for others for batching, which is better fit for the online API service.
• the currently startup time is ~20-30s, and inference speed is ~6s per minute of audio.
• if you need to access the codes of vLLM implementation, use this specific commit: https://github.com/SunbirdAI/worker-vllm/blob/718965fbf387298f01a7765023c67ea277268c92/modal-deploy/asr-whisper-large-v3-salt/batched_whisper.py

Usage

Modal dashboard: https://modal.com/apps/sb-modal-ws/main/deployed/asr-whisper-large-v3-salt

Deploy the asr-whisper-large-v3-salt model with Modal:

modal deploy batched_whisper.py

And query the endpoint with:

python client.py \
  --url https://sb-modal-ws--asr-whisper-large-v3-salt-model-transcribe.modal.run \
  --audio "../../sunflower-ultravox-vllm/audios/context_eng_6.wav"