Fix: SpeechToTextTool returns incoherent transcripts

[Martin-Labenne]

This PR improves SpeechToTextTool to ensure it works out of the box as expected, addressing issue #1478.

At a high level:

• Explicitly pass the correct whisper sampling_rate to the pre-processor and ensure the attention_mask is properly generated then passed through to the model’s forward method to avoid unreliable behavior during inference.
• Introduces explicit language control when instantiating the class (ex. language='en') while preserving the option for auto-detection when desired (language is omitted or None).
• Ensures audio inputs are resampled to the expected model sampling rate to avoid silent transcription failures.
• Adds support for longer audio files to to get transcription completeness.

Note: This is my first contribution to an open-source project, I'm excited to be part of it and open to any feedback or suggestions. Thanks a lot!