Pydgn

Changed

- (Needs simple test) Setting ``CUDA_VISIBLE_DEVICES`` variable before cuda is initialized, so that in ``--debug`` mode we can use the GPU with the least amount of used memory.
- Commented a couple of lines which forces OMP_NUM_THREADS to 1 and Pytorch threads to 1 as well. It seems we don't need them anymore.

Fixed

- to comply with `TUDataset`, we do not override the method `__len__` anymore

Fixed

- `load_dataset` does not assume anymore that there exists a `processed` data folder, but it is backward-compatible with previous versions.
- fixed an indexing bug on target data for node classification experiments (caused program to crash)
- Metric: renamed `_handle_reduction` to `_expand_reduction` and created a new helper routine `_update_num_samples` to allow a user to decide how to compute and average scores.

Fixed

- use of fractions of GPUs for a single task
- changed signature in forward to allow a dictionary (for MultiScore) or a value (for basic metrics)
- added squeeze in MulticlassAccuracy when target tensor has shape (?, 1)

Fixed

- Same bug as before but for `pydgn-dataset` =).

Fixed

- Bug that prevented locating classes via dotted paths in external projects.