Gaussian blur optimization [PyTorch backend]

[jamesdolezal]

Feature

The Gaussian blur augmentation in the PyTorch backend differs slightly from the Tensorflow implementation, as it uses torchvision.transforms.GaussianBlur, requiring odd-numbered kernel sizes only. It is also much slower, with throughput on a 3960X / A100 system of 500 img/sec (compared to the Tensorflow implementation at 7000 img/sec).

Pitch

Optimization of the Gaussian Blur augmentation to improve image throughput.