whisper transcription compatibility with newer transformers

[Saganaki22]

• Fix WhisperConfig max_length AttributeError — Newer versions of the transformers library removed the max_length attribute from WhisperConfig, causing transcription to crash with AttributeError: 'WhisperConfig' object has no attribute 'max_length'. Changed to use getattr() with a fallback so it works on both old and new versions of transformers.
• Fix float16 dtype mismatch — When running Whisper in half precision (float16), the input_features tensor was being passed as float32, causing a RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same. Now the input features are cast to match the model's dtype before inference.