Smarter way to count branches

[gordonwatts]

• Move to using count_nonzero one per axis to reduce # of tasks.
• This requires some iterative work for the counting rather than the straight forward signle line from before.

On a 2 file run this goes from 2230 tasks to 2228. Hmmm....

Fixes #48

[alexander-held]

After changes I would expect the numbers reported in the dashboard to be something like the number of files times the average amount of steps taken in the file (something like a factor 2 or so in the materialize_branches notebook I believe). I saw a very drastic reduction of almost two orders of magnitude. Examples for the graphs before and after are in https://github.com/iris-hep/idap-200gbps/pull/7, which shows the CMS version but that behaves the same. How does .visualize(optimize_graph=True) on the graph look like in this case here?

[gordonwatts]

Ok - here is before @alexander-held optimization trick:

And after:

[gordonwatts]

So, this is doing what we expect. The reason I was fooled was because I was doing len(total_count.dask) and that:

Before Optimization:

0003.5013 - INFO - Number of tasks in the dask graph: 172

After Optimization:

0003.5437 - INFO - Number of tasks in the dask graph: 118

So that looks like it should be 15, not 118. In short - I do not understand what len(total_count.dask) is doing.

[gordonwatts]

See issue #65 for follow up for the counting number (optimized graph vs non-optimized?).