Dataset can't be pickled

[zhiltsov-max]

Pickling can be useful for dumping and loading and for multiprocessing. It is required for PyTorch, because it heavily relies on multiprocessing in their Dataset. Currently, in many cases Datumaro will fail to pickle, mostly because of lambdas and local declarations.

> pickle.dumps(dataset)
E AttributeError: Can't pickle local object 'Environment.__init__.<locals>._filter.<locals>.<lambda>'

There are some places where such errors arise.