How many rows of theta to run through the MLP/Cholesky assembly at once

Mirrors the theta-blocking cross_iid() does for the flow estimators, so an MDN bounds the (theta, x) pair count the same way: mdn_mixture() materializes a (theta_chunk, K, dim_theta, dim_theta) tensor for one block, and theta_chunk * n_obs <= max_batch keeps that bounded regardless of which of theta or x is the large dimension (#240).

Usage

mdn_theta_chunk_size(n_theta, n_obs, max_batch)