Add decompression preprocessing step to TotalSegmentator2D for more efficient slice loading

[nkaenzig]

Closes #704

The issue for the slow data loading was due to the .gz compression used for the ct and mask .nii.gz files combined with reading individual slices. While reading the first few slices from compressed niftis using nibabel is fast, the deeper the slice index, the slower it gets, because each time you read a slice a sequential decompression of the file occurs. When unpacking the .nii.gz file beforehand, and then reading the .nii, reading deeper slices becomes much faster.
This is especially an issue for reading the data in 2D fashion, because if you do 3D you might read the whole CT only once, so you have to decompress only once.

This PR reduces the runtime for iterating over the complete dataset from 1h to 2-3min, using a torch dataloader with 16 workers.