…to categorical variables. Replaced the threshold attribute of SplitRecord and Node with SplitValue.

…hat defaults to -1 for each feature (indicating non-categorical feature).

…ly does nothing.

…(). Enforced max n_categories=2**31 for RandomSplitter.

…rical features.

…rating a big list of random numbers for every sample in every node when sending it left or right.

…hat there are no sparse categorical variables, since they're currently not supported.

…olds.

…ith more than 64 categories.

… to make sure we free the bit cache.

…ight

…splits until a non-trivial one is found

…rectly with ensembles. Refactored the categorical transformations into functions.

…identally creates a trivial split.

…resented in the local sample.

…gorical features.

…iate for categorical features.

…n with numpy v1.6. This should be reverted when support for numpy v1.6 is dropped.

… of numpy.

…ees, forests, and gradient boosting. Tweaked unit tests to match.