Support for unordered categorical features

[andrewherren]

Supporting unordered categorical features requires the following modifications:

• TrainData class must keep track of which features are categorical (ideally encoding them with integers for space efficiency)
• Tree class must be able to represent set-defined categorical splits
• XBART split sampler cannot generally enumerate all possible categorical splits, so should use the Fisher (1958) approach
• BART split sampler should use the Deshpande (2023) approach