chore: test xfail unification

[MarcoGorelli]

We're xfailing some tests in a variety of ways. I think it would be good to standardise somewhat

• requests.applymarkers(pytest.mark.xfail): this currently fails, but we aim to be able to support it in the future
• pytest.mark.skipif: there's some condition under which this just can't be tested (e.g. pyarrow types in pandas for pandas versions before they were supported)
• pytest.raises: this raises, and we expect it to raise

We haven't been super-consistent about this internally. I think it would be good to:

• document these standards in contributing.md
• apply them as much as possible internally. This would require looking over the current test suite and making some judgement calls - feel free to ask on Discord about anything you're not sure about!