Improve efficiency of `__getitem__`

[MarcAntoineSchmidtQC]

Currently, our approach for some of the __getitem__ methods is inefficient. For example, column subsetting for CategoricalMatrix converts the full matrix to a csc_matrix.

Here's a list to update with potential improvements:

• DenseMatrix: nothing to do. Already optimized with np.ndarray
• SparseMatrix: nothing to do. Already optimized with sps.csc_matrix
• CategoricalMatrix:
  • row: nothing to do, trivial
  • column: create a SparseMatrix with only the subset of columns/rows selected
• SplitMatrix:
  • Test thoroughly all the potential ways to index
• StandardizedMatrix
  • Not sure if columns subset with only one row works
• Write docstrings for expected behavior
• Write tests covering all expected behavior