fix: pandas and double nested vectors issue 885

[ioanaif]

Double nested ROOT arrays when opened with pandas their type remained to be an STLVector. This happened because AwkwardForth was not being invoked for pandas and the processing result was a numpy array which was then added to a series without any other transformations.

[jpivarski]

I think this is generally the right solution: there are array types that uproot.interpretation.library.Pandas.finalize should handle, but doesn't.

However, passing any array through ak.to_numpy(array) is too all-inclusive. Also, using any exception in that (and also the pandas.Series constructor) to trigger awkward-pandas is also too broad. There are details below: the safest way to fix this particular bug is to catch the previously-unhandled type of array (is it ObjectArray?) and prepare that the same way Awkward-handling code does, then pass it to awkward-pandas.

(I have a suggested change below, but the more I think about it, that suggested change is wrong: it assumes that the previously-unhandled type of array is Awkward, but I think it might be ObjectArray.)

[jpivarski]

If array is a NumPy array, we don't want to convert it into an Awkward Array just to convert it back.

Suggested change

	try:
	return pandas.Series(uproot.extras.awkward().to_numpy(array), index=index)
	except:
	if (
	isinstance(array, numpy.ndarray)
	and array.dtype.names is None
	and len(array.shape) == 1
	):
	return pandas.Series(array, index=index)
	elif type(array).__module__.startswith("awkward"):
	try:
	tmp = array.to_numpy()
	except ValueError:
	pass
	return pandas.Series(tmp, index=index)
	if True:

The next section is not appropriately indented, so I included a temporary if True: that will allow it to be tested, but it should be fixed.

Generally, we want try-except logic to be as narrow as possible because over-catching exceptions hides bugs.

[jpivarski]

In the originally failing case, what is the actual type of array? The safest way to fix it would be to only add a check for that additional type. It might mean we'd have to come back to it if there's another case we didn't handle, but fixing them one at a time minimizes the chance of breaking currently-working code.

If array is an ObjectArray because it went through the no-Forth fallback, then there are other steps in preparing the Python objects it contains as an Awkward Array (to wrap in a Pandas column). Specifically,

uproot5/src/uproot/interpretation/library.py

Lines 570 to 582 in fcb7b14

	elif isinstance(interpretation, uproot.interpretation.objects.AsObjects):
	form = json.loads(
	interpretation.awkward_form(interpretation.branch.file).to_json()
	)
	unlabeled = awkward.from_iter(
	(_object_to_awkward_json(form, x) for x in array), highlevel=False
	)
	return _awkward_add_doc(
	awkward,
	awkward.Array(_awkward_json_to_array(awkward, form, unlabeled)),
	branch,
	ak_add_doc,
	)

The intermediate steps are needed to set numeric types appropriately (Python floating point numbers are float64, but the Form might call for float32) and it adds doc strings to the Awkward __doc__ parameter (so that it shows up in Jupyter and IDEs).

(The best fix is likely different from my suggestion above.)

[jpivarski]

Perfect! Thank you!

I think it's done and can be merged, but I'll let you do that in case you have anything to add at the last minute.