MartinuzziFrancesco/torchrecurrent

AI Summary: Implement a new recurrent neural network layer, UGRNN, based on the provided equations and research paper. The implementation should be compatible with PyTorch and integrate seamlessly with the existing `torchrecurrent` library.

AI Summary: Rename the `activation_fn` keyword argument to `nonlinearity` in the `torchrecurrent` library's recurrent cell implementations to align with PyTorch's naming conventions. This involves modifying multiple files to update the argument name and potentially handling any downstream effects of the change.

AI Summary: Add individual bias toggles (e.g., `recurrent_bias`, `normal_bias`) to each recurrent layer implementation within the `torchrecurrent` project, allowing users to selectively enable or disable specific bias terms instead of controlling all biases with a single flag.