Document that if using n_chunks to trade off memory for speed, also should set cache_likelihood=False

[edeno]

Otherwise the likelihood will be computed and stored for the entire dataset and this usually will be too much for users that are running out of the GPU memory.

[edeno]

Another thing that may help, set the following environmental variable to use all of the available GPU memory:

os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "1.00"