Investigate GPU-Accelerated Image Preprocessing

@NetZissou is already working on this.

Since Jul 16, 2025.

# Overview
Currently, image preprocessing (such as resize and normalization) is performed on the CPU, which can become a bottleneck for high-throughput pipelines. Some libraries allow moving these steps onto the GPU, which can dramatically speed up batch inference if images are already loaded as tensors or in GPU memory.

# Relevant Libraries

- [torchvision.transforms.v2](https://pytorch.org/vision/stable/transforms.html#transforms-on-torch-tensors)
- [Kornia](https://kornia.org/)
- [NVIDIA DALI](https://docs.nvidia.com/deeplearning/dali/user-guide/docs/index.html)

# Potential Benefits

- Reduce CPU-GPU bottleneck during image preprocessing
- Increase throughput for large-scale inference tasks

# Action Items

- Explore feasibility of moving preprocessing steps onto the GPU using the above libraries.
- Benchmark performance vs. current CPU-based approach.
- Assess ease of integration with the current pipeline