Other things that need fixing

• Remove 256 large stack allocations for left/right
• Add a way to monitor the training process. Probably on a per-tree level. Need a train_step type C function. Maybe a generator on Python side, as a convenience function
• Fix hardcoded stack_size
• Actually use randomized threshold selection. Right now there is logic more akin to RandomTrees, but with hardcoded 50 unique values (seemingly computed without random sampling). Goes through all samples - probably a major speedup if/when fixed.
• Missing check on n_samples wrt max_samples in train
• References to sample indices should be uint16_t, not int16_t
• Types for general integers should be int, not int16_t
• Test that the random number generator works reasonably