Inconsistent inference acceleration parameters across functions + additional optimization opportunities

[Mr-Milk]

Several functions in LazySlide accept parameters for accelerating model inference (amp, autocast_dtype, compile, compile_kws), but these are not consistently available across the full inference pipeline. Additionally, there are other well-established PyTorch acceleration techniques that could be integrated.

• tl.feature_extraction: Missing compile, compile_kws
• tl.feature_aggregation: Missing compile, compile_kws
• tl.tile_prediction: Missing compile, compile_kws
• tl.text_embedding: Missing compile, compile_kws
• tl.zero_shot_score: Missing amp, autocast_dtype, compile, compile_kws
• tl.slide_caption: Missing amp, autocast_dtype, compile, compile_kws
• tl.virtual_stain: Missing compile, compile_kws
• tl.image_generation: Missing compile, compile_kws
• seg.cells: Missing compile, compile_kws
• seg.cell_types: Missing compile, compile_kws
• seg.tissue: Missing compile, compile_kws
• seg.artifact: Missing compile, compile_kws

Dataloader optimizations:

1. Set pin_meory=True?
2. Expose prefetch_factor to the user?
3. Set persistent_workers=True when num_workers > 0