Torchvtk

Transforms have changed so that when using a Transform with `apply_on=['a','b']`, the same transform is applied to both `a` and `b`.
With this you can for example do a

RandPermute(apply_on=['input', 'label'])

to apply the same random permutation to both an input and its label (e.g. volume and segmentation volume).


Also added `tqdm` to dependencies.

This changes behavior of tiling datasets (`TorchDataset.tile()`) to NOT include a remainder of a volume when tiling (with bigger overlap), but instead leave the remainder untouched and shift the tile locations to the center. So if some data shape cannot be tiled without remainder, the remainder will be at the border of the data. The maximum amount omitted along one dimension is `shape[i] - overlap[i] -1`, half of it in the beginning and half of at the end of the given dimension `i`.

TorchDataset` now has a `.tile()` method that automatically splits the data into tiles with a given tile size and overlap.
This also works together with `.preload`.
A tiled dataset will compute all tiling locations once in the beginning, if preloaded, otherwise on the fly. That means that a non-preloaded dataset will just return a random tile and does not have adjusted `__len__` to loop once over all tiles in the dataset. The tiling further adds a `'num_tiles'` and `'tile_location'` entry in the loaded tile dictionary to preserve the location information. This information is loaded directly in your full data dictionary if the dataset is preloaded.

neat import errors