Reduce data embedding in cohorts graph

[dcherian]

flox/flox/core.py

Lines 1949 to 1961 in b20545b

	cohort_index = pd.Index(cohort)
	reindexer = (
	partial(
	reindex_intermediates,
	agg=agg,
	unique_groups=cohort_index,
	array_type=reindex.array_type,
	)
	if do_simple_combine
	else identity
	)
	subset = subset_to_blocks(intermediate, blks, block_shape, reindexer, chunks_as_array)
	dsk |= subset.layer # type: ignore[operator]

Many copies of cohort_index are embedded in the graph. Instead, we should just embed it under its own key, and then add references to the graph layer.