Support reindexing in simple_combine

[dcherian]

• add "grouped" or "simple" in task name and check layer names to make sure we don't regress

This PR allows reindexing inside _simple_combine so that we don't need to rerun
the groupby-reduce at every combination stage.

This means we are down to three options (good improvement!)

1. map-reduce with reindex=True: reindex to expected_groups at blockwise stage, then map-reduce
2. map-reduce with reindex=False: reindex at combine stage of map-reduce
3. cohorts : start with reindex=False at blockwise, and then reindex to cohorts specifically, then tree reduce with reindex=True

For 1D combine, we see great improvements for reindex=True
reindex=False uses 2x more memory but similar time.

Note that the reindex=False intermediates are a worst case where there are
no shared groups between the chunks being combined.
This case is actually optimized in _group_combine where reindexing is
skipped for reducing along a single axis.

[ 68.75%] ··· =========== ========= =========
              --                combine
              ----------- -------------------
                  kind     grouped   combine
              =========== ========= =========
                reindexed      760M      631M
               not-reindexed    981M     1.81G
              =========== ========= =========

[ 75.00%] ··· =========== ========== ===========
              --                 combine
              ----------- ----------------------
                  kind     grouped     combine
              =========== ========== ===========
                reindexed    393±10ms    137±10ms
               not-reindexed   652±10ms   611±400ms
              =========== ========== ===========

[dcherian]

Looks like reindex=False is a major improvement when worker-saturation=inf but when worker-saturation=1.2, then it's all basically the same :(