Target store returned from StoreToZarr without waiting on completion of StoreDatasetFragments

[cisaacstern]

This is my mistake. When I merged #562, I did not understand that Python line numbers were not used for determining sequencing of the Beam DAG. I now see that by not chaining the returned target_store onto the output PCollection of StoreDatasetFragments here

pangeo-forge-recipes/pangeo_forge_recipes/transforms.py

Lines 484 to 485 in 2739f22

	rechunked_datasets | StoreDatasetFragments(target_store=target_store)
	return target_store

Beam optimizes by returning target_store before StoreDatasetFragments is complete. AFAICT the solution discussed in #556 (comment) makes the most sense. To summarize:

1. return target_store from StoreDatasetFragments so that it emits a PCollection of length n target_stores
2. take a fixed size beam.Sample from the target stores PCollection so we get a singleton PCollection containing a single target_store
3. return that singleton target_store

xref leap-stc/cmip6-leap-feedstock#9 (comment)

[cisaacstern]

I can work on this Monday