feat: update A2T audio conversion

[ad-astra-video]

PR improves file handling of audio-to-text pipeline by removing call to ffmpeg to process the audio and uses pyav to convert to raw audio.

Cloud SPE test file fails in my testing sometimes with current conversion implementation using "mp3" container and format. The processing error was coming from the internal conversion to raw audio in transformers (links below). Switching to processing the audio file using pyav allows the Cloud SPE file to process correctly.

Transformers calling ffmpeg binary from preprocess function.
https://github.com/huggingface/transformers/blob/47c29ccfaf56947d845971a439cbe75a764b63d7/src/transformers/pipelines/automatic_speech_recognition.py#L353
https://github.com/huggingface/transformers/blob/47c29ccfaf56947d845971a439cbe75a764b63d7/src/transformers/pipelines/audio_utils.py#L10

Some marginal speed improvements:
Cloud SPE test file (3s clip)
965ms - bytes of container sent to model
895ms - np ndarray sent to model

Another test file (3m 22s)
5.6s - bytes of container sent to model
5.5s - np ndarray sent to model

[eliteprox]

LGTM! Tested with MP4 and regression MP3 and FLAC. Left a few nits for cleanup

[ad-astra-video]

I made updates for small clean up and removed import not needed. Rebuilt docker container and tested to confirm both test files work (one test file is clouds test audio file).