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"] = ".99"

Need to do this before importing jax