Switched model architecture for the text detection and layout models:

• 30% faster on GPU
• 4x faster on CPU
• 12x faster on MPS (M series macs)

Accuracy should be about the same, or slightly better, from my benchmarks.