Start migrating I/O writers to pylibcudf (starting with JSON)

[lithomas1]

Description

Switches the JSON writer to use pylibcudf.
xref #15162

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wence-]

Thanks, I think there a few organisational tidy-ups we can do in the tests.

[wence-]

Wah?! This is "once-only" fixture, no, so it should be cleaned up automatically. What is this for?

[lithomas1]

I couldn't figure out how to reset the StringIO/BytesIO buffers 😅 .
(otherwise we have the writes from the previous calls to write_json in it)

I tried setting the scope to "function", but that didn't seem to do anything.

[wence-]

It should be the case that pytest conses a new object every time this fixture is used, so it seems something is wrong :(

[wence-]

Can you distill this to a smaller test case, because this sounds absolutely like a pytest bug.

[lithomas1]

I think I get it now.
The issue is that the StringIO and the BytesIO are created outside of the fixture (since they are passed in the params list).

Every time pytest runs a test, it probably runs the fixture again (passing the same StringIO/BytesIO instance from the params list into the fixture, but since we return the params unchanged, we end up reusing the StringIO/BytesIO).

(This is why calling copy.copy on fp_or_buf inside of the fixture - which I just tested - would also fix the issue.)

I guess the question in that case would be should pytest make a copy of the input parameters.
(probably not by default at the very least, I assume there's some classes that disallow copying)
Even if pytest doesn't change anything, this is still probably worth reporting so that they can add it to the docs or something.

[wence-]

Ah ok, yes, I was blind to this. This is just standard python, because pytest only sees the actual objects that have already been constructed (it's the same trap as def foo(param=[]))

The right way to handle this is something like:

@pytest.fixture(params=["string", "pathlib.Path", "BytesIO", "StringIO"])
def source_or_sink(request, tmp_path):
    if request.param == "string":
        return f"{tmp_path}/{request.param}"
    elif request.param == "pathlib.Path":
        return pathlib.Path(tmp_path) / request.param
    elif request.param == "BytesIO":
        return io.BytesIO()
    elif request.param == "StringIO":
        return io.StringIO()
    assert False

[lithomas1]

Pushed something up like this.

[lithomas1]

All skips in test_copying should be now removed, thanks Lawrence for the help!

[wence-]

I don't have any more blocking comments I think. Thanks. I will let @vyasr decide if there are any more from their point of view.

My worry about needing to reset the StringIO bufs in the test fixture still applies. I absolutely don't think that should be necessary.

[vyasr]

Couple of small things left, but enough that I think it's worth fixing before we merge

[vyasr]

Did this PR get opened? If not, can we either make it, or at least open an issue so that we don't forget?

[lithomas1]

OK, switched it back to loops in conditionals.

[vyasr]

Couple of last suggestions, but I'm approving now so feel free to merge once you've addressed as you see fit.

[lithomas1]

/merge