Upgrade to DataFusion 43, fix a bug, add more tests

[andygrove]

This is based on #50 and adds an extra test and fixes a bug

[edmondop]

Can you explain why in the case of a leaf node, the input partition is the same as the output partition of the plan, while in case of a plan with children it is the output partitioning of the first child? This is assuming that all children have the same partition count?

[andygrove]

in context.py we have this logic:

    # if the query stage has a single output partition then we need to execute for the output
    # partition, otherwise we need to execute in parallel for each input partition
    concurrency = stage.get_input_partition_count()
    output_partitions_count = stage.get_output_partition_count()

This is based on the assumption that the query stage is a shuffle write, which perhaps was always true when running TPC-H, so the existing code worked.

With the new simple SELECT * FROM table test that you added, we had a query stage where the plan was a CsvExec and had no children so we had to handle this as a special case. There is no input partitioning in this case. We use the output partitioning because DataFusion will have already decided that based on the files that are available.

This code is all confusing and I would like to make it less so.