Update Phi model

[makaveli10]

• Use Phi finetuned model with chatml prompt format.
• Always torch.compile WhisperSpeech for faster inference.