wip: fix axis=0 concatenation

[agoose77]

Fixes #456

It's clear from #456 that the complexity around axis=0 concatenation makes it hard to reason about!

The bug is caused by the loss of report objects associated with secondary arrays in the concatenate call; right now we take the first array (because all arrays are coerced to the same form) as the "meta".

To fix this, we need to build a "parent" meta such that parent.touch_data("key-1") invokes child1.touch_data("key-1-child"), child2.touch_data("key-2-child"), and so on.

I'll try and clean this up, probably with an Awkward fix, in the near future.

[codecov-commenter]

Codecov Report

Attention: 17 lines in your changes are missing coverage. Please review.

Comparison is base (d3b6208) 93.10% compared to head (c62796d) 92.72%.

Files	Patch %	Lines
src/dask_awkward/lib/operations.py	73.01%	17 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #457      +/-   ##
==========================================
- Coverage   93.10%   92.72%   -0.38%     
==========================================
  Files          23       23              
  Lines        3293     3355      +62     
==========================================
+ Hits         3066     3111      +45     
- Misses        227      244      +17     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jpivarski]

Is this affected, or even made easier, by scikit-hep/awkward#3043 (and subsequent changes to dask-awkward to defensively copy reports instead of changing them in place throughout a DAG)?

Once the reports become effectively immutable from dask-awkward's point of view—that is, each DAG node is associated with an unchanging report—then we won't need to have a hierarchical structure, nesting reports, since we could make the output of dak.concatenate propagate all the reports from its inputs.

I came to this from #456, since @gordonwatts hit the same issue again and I was checking up on it.

[agoose77]

I think this will become slightly simpler, but much of the logic will remain. The feature required by result = ak.concatenate((part1, part2, ..., partN)) is that touching result.pt.data also touches part1.pt.data, part2.pt.data, ..., partN.pt.data. I think we'll need some kind of hierarchy. We could encode that as a hierarchical report or with a parentage attribute in the report.