Speed up legacy polygon extraction

[anutkk]

Speed up legacy polygon extraction using light software optimizations:

• use np.asarray instead of np.array to convert PIL image to Numpy array
• use PIL.ImageDraw.Draw instead of skimage.draw.polygon to create the polygon mask
• use a faster, vectorized version of skimage.transform.PiecewiseAffineTransform instead of vanilla PiecewiseAffineTransform

I tested this segmentation on the BiblIA Italian subset (3748 lines). Results:

Method	Accuracy	Runtime (minutes:seconds)
Current legacy extraction	97.29%	56:27
Slightly optimized legacy extraction	97.33%	29:26

The speedup is more pronounced for straight baselines (closer to 4x).

The difference in accuracy is only 0.04%.

[anutkk]

Cancelling due to typo