Introduce an interface for loading and preprocessing data

[juliabeliaeva]

This PR introduces a DataLoader interface which represents the abstraction for loading data in a form of FloatArray with a TensorShape from some source, e.g. a file. Currently this is done in one step: loading and preprocessing together and there is no access to the intermediate data (e.g. BufferedImage). So far this seems OK to me, but we may want to change it in the future. Currently the interface is mainly used in the datasets, but I intend to use it for onnx models to make them more platform-independent (we should have another implementation which uses Bitmap as a data source for android).

After these changes OnFlyImageDataset needs a new name since it can work with any data now, not just images, and also "on fly" is not grammatically correct (should be "on the fly", and it also seems more like an adverb, not an adjective). My current candidates are LazyDataset, LoadingDataset and LightweightDataset.

Since DataLoader uses TensorShape, the latter had to be moved to the "dataset" module and it does not look too good. Perhaps, "api" module should not depend on the "dataset" and it should be the other way around.

[juliabeliaeva]

I removed Preprocessing implementing DataLoader<File> from this branch so that not to interfere with the #425. We can add it back later if we want to when the work in #425 would be finished. Right now I use dataLoader() extension function to convert Preprocessing to a DataLoader which seems to be enough at the moment.
Also it may be a good idea to move loading of BufferedImage from file out of Preprocessing.