GPU accelerated batched pre/post processing

[MarcelLieb]

What does this change?

• Reimplement pre/post processing using torchvision functions, which allows processing multiple images simultaneously on the GPU (Post-processing adapted from https://github.com/99oblivius/CorridorKey-Engine)
• add toggle to select between opencv and pytorch post-processing pipeline
• add tests for both torch and opencv methods
• add tests for batched processing
• add ability to opt out of composite previews generation
• implement higher quality despeckle
• total processing time reduction ~30% (potential of up to 2 fps on 5070 Ti if image loading gets optimized)
• change the model to full float16 precision

Differences of main branch vs pre-processing on GPU

Differences of main branch vs post-processing on GPU

Differences main branch vs full GPU pipeline

Differences full GPU fp32 vs fp16

Improved despeckle

Missing

To make full use of these optimizations clip_manager needs to be reworked to allow batch processing.
Also these changes need to be implemented into the mlx model to avoid a large divergence between implementations.

Checklist

• uv run pytest passes
• uv run ruff check passes
• uv run ruff format --check passes