Integrate training from sbb pixelwise segmentation

[kba]

This merges the unifying-training-models branch of https://github.com/qurator-spk/sbb_pixelwise_segmentation into the folder train, so we have everything needed for running and training eynollah's components in a single place.

I did the merge with:

cd ../sbb_pixelwise_segmentation
git filter-repo --force --to-subdirectory-filter train
cd ../eynollah
git add remote train ../sbb_pixelwise_segmentation
git merge --allow-unrelated-histories -m "Merge sbb_pixelwise_segmentation training code into eynollah" train/unifying-training-models

So, all the commits of sbb_pixelwise_segmentation are included in eynollah's git history now.

I struggled with the documentation, though. The documentation on training models was in different states in three places AFAICS: sbb_pixelwise_segmentation.wiki, train.md in sbb_pixelwise_segmentation and in docs/train.md in eynollah. I did align the texts syntactically to compare and decided that docs/train.md is the most up-to-date version, so I removed the others. While resolving the train.md version, I also added the eynollah wiki page documenting (an earlier stage of) eynollah (layout) as docs/eynollah_layout.md. Unless there are good reasons against, I would recommend closing that wiki and keeping all documentation in /docs.

Once we merge here, I would:

• transfer the remaining issues in sbb_pixelwise_segmentation to eynollah
• Close or rebase the remaining PR (@vahidrezanezhad please review Optionally Reduce LR on plateau sbb_pixelwise_segmentation#25 and tf.keras version that allows any input resolution and doesn't use Lambda layers sbb_pixelwise_segmentation#16)
• Archive sbb_pixelwise_segmentation and update documentation to point to eynollah

[cneud]

All good to go now, thx!