Add category matrix support to AttributeSet

As discussed today with @sushobhan2024, we need to add a new cat_matrix method to AttributeSet and its subclasses, to compute a matrix over categorical values for various attribute representations.

• Define the cat_matrix signature. Proposed:

  def cat_matrix(self, *, normalize: Literal["unit" | "distribution"] = None) -> NDArray[np.floating[Any]] | csr_array: ...
  

• Implement cat_matrix for scalar attribute sets, where the columns of the matrix refer to distinct values of the attribute. It should throw a TypeError if the attribute is floating-point. We may want to cache a Vocabulary of distinct values to speed up repeated calls. If the attribute is null for a given item, the resulting matrix should be all 0. Should usually return a sparse matrix (csr_array).

• Implement cat_matrix for list attribute sets. Behaves like scalar attribute, except that more than one value can be nonzero.

• Implement cat_matrix for dense vector attribute sets. Returns a dense matrix (np.ndarray), columns correspond to the attribute vector columns. Basically just returns the vector as-is, possibly with normalized rows.

• Implement cat_matrix for sparse vector attribute sets. Same as dense, except it returns a sparse matrix instead of a dense one.

Open question: do we provide access to the vocabulary for list and scalar attribute sets, and how? Maybe we also add a cat_vocabulary() method that returns the column vocabulary for a vector attribute set, and the unique value vocabulary for a scalar or list set, and None if the scalar or list is floating-point or if the vector attribute has no dimension vocabulary / names.