Pybnesian

- Fixed a bug in `DiscreteFactor` and others hybrid factors, such as `CLinearGaussianCPD` and `HCKDE`, where categorical data would not be correctly validated. This could lead to erroneous results or undefined behavior (often leading to segmentation fault). Thanks to Carlos Li for reporting this bug.

- Support for Python 3.10 and `pyarrow>=9.0` has been added. Support for Python 3.6 has been deprecated, as `pyarrow` no longer supports it.

- manylinux2014 wheels are now used instead of manylinux2010, since `pyarrow` no longer provides manylinux2010 wheels.

- Fixed important bug in OpenCL for NVIDIA GPUs, as they define small OpenCL constant memory. See [https://stackoverflow.com/questions/63080816/opencl-small-constant-memory-size-on-nvidia-gpu](https://stackoverflow.com/questions/63080816/opencl-small-constant-memory-size-on-nvidia-gpu).

- Added support for Apache Arrow 7.0.0.

- Added method `ConditionalBayesianNetworkBase.interface_arcs()`.
- `GreedyHillClimbing` and `MMHC` now accepts a blacklist of `FactorType`.
- `BayesianNetworkType.data_default_node_type()` now returns a list of `FactorType` indicating the priority of each `FactorType` for each data type.
- `BayesianNetworkBase.set_unknown_node_types()` now accepts an argument of `FactorType` blacklist.
- Change `HeterogeneousBN` constructor and `HeterogeneousBNType.default_node_types()` to accept lists of default
`FactorType`.
- Adds constructors for `HeterogeneousBN` and `CLGNetwork` that can set the `FactorType` for each node.

- Bug Fixes:

- An overflow error in `ChiSquare` hypothesis test was raised when the statistic were close to 0.
- Arc blacklists/whitelists with repeated arcs were not correctly processed.
- Fixed an error in the use of the patience parameter. Previously, the algorithm was executed as with a `patience - 1` value.
- Improve the validation of objects returned from Python class extensions, so it errors when the extensions are not correctly implemented.
- Fixed many serialization bugs. In particular, there were multiple bugs related with the serialization of models with Python extensions.
- Included a fix for the Windows build (by setting a correct `__cplusplus` value).
- Fixed a bug in `LinearGaussianCPD.fit()` with 2 parents. In some cases, it was detecting a linear dependence between the parents that did not exist.
- Fixes a bug which causes that the Python-class extension functionality is removed.
Related to: [https://github.com/pybind/pybind11/issues/1333](https://github.com/pybind/pybind11/issues/1333).

- Improvements on the code that checks that a matrix is positive definite.
- A bug affecting the learning of conditional Bayesian networks with `MMHC` has been fixed. This bug also affected `DMMHC`.
- Fixed a bug that affected the type of the parameter `bn_type` of `MMHC.estimate()`, `MMHC.estimate_conditional()` and `DMMHC.estimate()`.

- Adds support for pyarrow 5.0.0 in the PyPi wheels.
- Added `Arguments.args()` to access the `args` and `kwargs` for a node.
- Added `BayesianNetworkBase.underlying_node_type()` to get the underlying node type of a node given some data.
- Improves the fitting of hybrid factors. Now, an specific discrete configuration can be left unfitted if the base continuous factor raises `SingularCovarianceData`.
- Improves the `LinearGaussianCPD` fit when the covariance matrix of the data is singular.
- Improves the `NormalReferenceRule`, `ScottsBandwidth`, and `UCV` estimation when the covariance of the data is singular.
- Fixes a bug loading an heterogeneous Bayesian network from a file.
- Introduces a check that a needed category exists in discrete data.
- `Assignment` now supports integer numbers converting them automatically to float.
- Fix a bug in `GreedyHillClimbing` that caused the return of Bayesian networks with `UnknownFactorType`.
- Reduces memory usage when fitting and printing an hybrid `Factor`.
- Fixes a precision bug in `GreedyHillClimbing`.
- Improves `CrossValidation` parameter checking.