max_batch counts (theta, x) pairs, so that is what sizes the chunk. The
quadratic form materializes K components and p dimensions per pair on top
of that, so a second cap keeps the (T, K, n, p) intermediate to a few tens
of megabytes however many components the mixture has. Dividing by K * p
alone, as this used to, chunked a 5000-observation call ten ways and paid the
fixed cost of a torch call ten times for a 400 KB intermediate.