Neuralogic

Breaking Changes:

- Removed DyNet and PyTorch backends support
- Removed `Backend` enum

Additions:

- New NeuraLogic as `torch.nn.Module`
- Gringo grounder experimental support
- New `next/2` special predicate
- Docs updates

Breaking changes:

- Removed `Optimizer` enum
- Optimizers are now located in `neuralogic.optim` and are implemented as classes
- Removed `learning_rate` from `Settings` - now part of optimizer parameters (e.g., `neuralogic.optim.Adam(lr=0.001)`)

Additions:

- Logarithm transformation function (`Transformation.LOG` or `neuralogic.nn.functional.log` (`F.log` with aliasing))

Bug fixes:

- Fixed broken matrix values as fact values

Breaking changes
- `Activation` renamed to `Transformation`
- `ActivationAgg` renamed to `Combination`
- `Combination` is now a separate field in metadata

Additions
- New combination and aggregation functions

Others
- Performance improvements

What's Changed
* fix order of parameters and mask for Data.from_pyg by joaquincabezas in https://github.com/LukasZahradnik/PyNeuraLogic/pull/35

New Contributors
* joaquincabezas made their first contribution in https://github.com/LukasZahradnik/PyNeuraLogic/pull/35

- Add `.T` sugar syntax for transposing body literals (e.g., `R.body_literal.T`)
- Add "functional" interface (`neuralogic.nn.functional`) for activation/aggregation functions

Additions:

- Compilation of (subset of) trained models into PostgreSQL (possible direct evaluation in database)

- `MetaConv` module
- `MAGNNMean` module (without attention)
- `MAGNNLinear` module (without attention)

- Recursive Neural Network (`RvNN`) module
- Recurrent Neural Network (`RNN`) module
- Long Short-Term Memory (`LSTM`) module
- Gated Recurrent Unit (`GRU`) module

- New `Activation` - `INVERSE` ($1 / x$), `REVERSE` $(1 - x)$ , `SQRT` ($\sqrt{x}$), `NORM`, `TRANSP`, `EXP`

- New dataset sources: `CSVDataset` (csv files/strings) and `DBDataset` (aimed for psycopg2 support, but other connectors for different dbs might work as well)

- `Template::remove_duplicates` method for removing duplicate rules
- `InferenceEngine::get_queries` method for listing all possible queries for a template

- Support for activation functions on body literals (e.g., `R.h(V.X) <= (Activation.SIGMOID(R.atom(V.Y)), R.edge(V.Y, V.X))`)

Changes:

- `Term` (`T`) renamed to `Constant` (`C`)
- Removed "adding" of `Activation` with `ActivationAgg` -> instead of `ActivationAgg.MAX + Activation.SIGMOID` use `ActivationAgg.MAX(Activation.SIGMOID)`
- `Activation`, `Aggreagation`, `ActivationAgg` are no longer enums
- Negation of literal (e.g., `-R.literal`) now applies `REVERSE` activation function on the literal (not the actual negation)

Bug Fixes:

- Many bug fixes