perf(whisper): lazy-import timing module to avoid loading scipy/numba

[YashD5291]

Summary

• Move from .timing import add_word_timestamps from module-level to inside the if word_timestamps: guard in transcribe()
• Makes scipy and numba truly optional when word_timestamps=False (the default)

Problem

timing.py imports scipy and numba at module level:

import numba           # timing.py:8
from scipy import signal  # timing.py:10

These are loaded unconditionally via transcribe.py's top-level from .timing import add_word_timestamps, even when word_timestamps=False. This eagerly loads ~620 Python modules and ~212 MB of native libraries that are never used.

Fix

One-line change: move the import inside the if word_timestamps: conditional where it's actually needed.

- from .timing import add_word_timestamps

  if word_timestamps:
+     from .timing import add_word_timestamps
+
      add_word_timestamps(...)

Impact

• Import time: import mlx_whisper no longer loads scipy/numba/llvmlite
• Packaging: Downstream apps using PyInstaller/cx_Freeze can exclude scipy (~71 MB), numba (~31 MB), and llvmlite (~110 MB) from bundles when word timestamps aren't needed
• Zero behavior change: When word_timestamps=True, the import fires on first use and everything works identically

Testing

• word_timestamps=False (default): scipy/numba never imported — verified via sys.modules inspection
• word_timestamps=True: import fires inside the conditional, all word timestamp functionality works as before