Fastfeedforward

- one no longer has to specify `use_hard_decisions=True` when running forward under `model.training=False` (`ValueError` will no longer be thrown)
- fixed a bug that occured for some PyTorch versions where torch.matmul broadcasting across two-dimensional batch size resulted in greedy memory allocation and unreasonable CUDA OOMs
- using `index_select` in `forward_eval` was fast but too memory-consuming; reduced it to a for-loop until `.foreach` or custom CUDA implementation is available

FFF
- renamed the `__init__` parameter `hidden_width` to `leaf_width` (yes, this disrespects semantic versioning a little, but hey ...)
- introduced the `region_leak` parameter
- introduced the `usage_mode` parameter
- added `get_node_param_group` and `get_leaf_param_group` methods to enable connection to `LocalSGD` and `LocalAdam` with ease
fastfeedforward.optim
- added as a separate package
- contains implementations of the `LocalSGD` and `LocalAdam`

This is the initial version with a basic implementation of the fast feedforward module and with an example notebook.