fix(audio): normalize negative axis in pad_or_trim

[loulanyue]

Summary

In whisper/audio.py:

• pad_or_trim accepts an axis parameter (default -1).
• When constructing pad_widths and accessing tensor/array dimensions, negative axis values are now explicitly normalized (dim = axis if axis >= 0 else array.ndim + axis).
• This ensures clean, consistent indexing across PyTorch tensor and NumPy array padding logic.