fix: complain about CannotBeAwkward earlier, before reading data.

[jpivarski]

This came up in Discussion #826 (Issue #833): the AwkwardForth errors occur on an Interpretation that could never have been Awkward anyway.

An early hint was that it was going through read_object_any, which means arbitrary-typed pointers, circular references, and a whole host of things that are beyond the scope of Awkward Array, but the clearest indication was that we got the "Cannot be Awkward" error in TTrees with zero entries—if it could have gotten past the AwkwardForth read, it would have encountered this error, anyway. (It didn't get past the AwkwardForth read because we don't generate AwkwardForth inside read_object_any.)

So this PR

• takes the check out of Library.finalize
• puts it into _ranges_or_baskets_to_arrays, on the main thread, before any reading starts¹
• fixes an unrelated error in _util, but it was uncovered by this PR.

Footnotes

1. Well, actually, Source.chunks has already been called, so the bytes have been read from local files or a remote server has been queried, but we haven't waited for any data to come back yet and won't be decompressing it or interpreting it if it does. ↩

[jpivarski]

This is unrelated to the main purpose of the PR, but calling the awkward_forth method at this point in the code revealed that the leaf-list case isn't handled. Somehow, test_0033-more-interpretations-2.py::test_leaflist_awkward was passing before because something must have figured out that it needs a RecordArray without calling awkward_forth. (I'm not sure which code path that is, but now, at least, this is consistent with that.)

[ioanaif]

LGTM! Should tests be added for this issue?

[jpivarski]

The file that revealed the issue is private—otherwise, that would have been used as a test.

However, every existing test with library="ak" goes through this new code path, so it is being tested in the sense of having full coverage. In fact, the unrelated bug (highlighted above) was discovered because it went through this code path, and in the old system, it somehow managed to avoid the old (late) code path. I'd take that as an indication that coverage has increased.