atc-model

Update .gitignore to exclude hf_token.txt in addition to whisper directory

Ensure input features match model dtype by converting to bf16 in DataCollatorSpeechSeq2Seq during evaluation.

Update training script configuration: adjust batch sizes, gradient accumulation, and add PyTorch CUDA allocation settings for improved memory management.

Set environment variables for thread management and improve CPU allocation in training script

Refactor parallel preprocessing to scale with available CPU cores