Neuralogic

First public release of the Lifted Relational Neural Networks concept (in development since 2013).

_Write simple declarative parameterized programs (templates) in the specified language (~Datalog) to encode advanced deep relational learning scenarios, and use the NeuraLogic engine (this jar file) to compile them and optimize their parameters on your relational data._

See examples for how to write the programs/templates. No warranties attached.

Breaking changes
- Remove PyG translation
- Remove `ModuleList`
- Switch settings and backend arguments (in `get_evaluator` etc.)
- Move inference engines from `neuralogic.core` to `neuralogic.inference`
- Move error functions from `neuralogic.core.error_function` to `neuralogic.nn.loss`
- Move intializers from `neuralogic.core.enums` to `neuralogic.nn.init`
- Move datasets from `neuralogic.core` to `neuralogic.dataset`
- Split `Dataset` into (Logic)`Dataset`, `FileDataset` and `TensorDataset`
- Rename settings parameters (`rule_neuron_activation` -> `rule_activation` and `relation_neuron_activation` -> `relation_activation`)

Additions
- Add random seed settings (`neuralogic.manual_seed`, `neuralogic.seed`, `neuralogic.initial_seed`)
- Add `neuralogic.nn.module.Linear`
- Add `neuralogic.nn.module.MLP`
- Add pooling modules (`neuralogic.nn.module.Pooling`, `neuralogic.nn.module.SumPooling`, etc.)
- Add `neuralogic.nn.module.ResGatedGraphConv`

Docs changes
- Add recursive xor documentation/example
- Document logging, debugging and Java settings
- Visual enhancements

- Add `parameters` as an alias for `state_dict`
- Fix matrix parameters serialization

- Newly introduced methods on templates, models, evaluators and samples for drawing graphs
- New predefined modules (SGConv, APPNPConv etc.)

- Usage of error function via `ErrorFunction` enum has been replaced with their own classes in `neuralogic.core.error_function` module.
- Setting the error function as `CrossEntropy()` or `SoftEntropy()` will set the activation function of the last layer to identity and apply the logsumexp trick.
- Setting the error function as `CrossEntropy(with_logits=False)` will set the activation function of the last layer to sigmoid/softmax and will apply standard CrossEntropy.
- Path to graphviz (dot) can now be parametrized via `graphviz_path` parameter of `draw_model`/`draw_sample` - e.g.:

draw_model(model, filename="model.png", graphviz_path="/usr/bin/dot")
draw_model(model, filename="model.png", graphviz_path="C:\xyz\dot.exe")
etc.

The dot executable is taken from the path (if present) by default.

- Introduction of the new modules (with added RGCN) that will serve as a replacement of modules in `utils.templates` in the future.

- Reimplemented Java backend logging
- Increased Java heap size
- Added JVM options setter
- Fixed reading output values from Java backend
- Fixed wrong imports in PyG
- Fixed the SquaredDiff error function