Testing: improve generic test to also test concrete expected values #1945

Right now the tests that every reader gets are the compliance checks in BaseTestRawIO.test_read_all. These are consistency checks (shapes, dtypes, different ways of reading the same data agreeing with each other) and as they are the same for every format they can't express what values a given file should contain.

However, these recent fixes were all for readers that returned the right shape with the wrong values:

• BrainVisionRawIO: read marker positions as 1-based #1933: BrainVisionRawIO used the 1-based marker positions as 0-based sample indices and every event came out one sample late.
• Fix BCI2000RawIO scaling: decimal gains were parsed as 0 and the offset was applied with the wrong convention #1932: BCI2000RawIO parsed decimal gains as 0 and every channel came out as a constant.
• Map EDF digital endpoints to their physical endpoints #1920: EDFRawIO computed the gain with a + 1 in the digital range.

What these PRs share is that each one adds a separate test against specific values which means we load the data again. I think we should include this mechanism within our testing framework. Each test class would list the values it expects from the files it already tests, and test_read_all would check them on the reader it has already parsed. This way we reuse the files we already download, we don't read them twice and we have one place to test values. A test class that lists nothing doesn't change:

class BaseTestRawIO:
    ...
    expected_values = {}

    def test_read_all(self):
        for entity_name in self.entities_to_test:
            ...  # build the reader, parse the header and run the compliance checks as today
            expected = self.expected_values.get(entity_name, {})

            event_names = list(reader.header["event_channels"]["name"])
            for event_name, timestamps in expected.get("event_timestamps", {}).items():
                actual, _, _ = reader.get_event_timestamps(event_channel_index=event_names.index(event_name))
                np.testing.assert_array_equal(actual, timestamps)

            if "scaled_samples" in expected:
                ...  # read the first samples, rescale and compare


class TestBrainVisionRawIO(BaseTestRawIO, unittest.TestCase):
    ...
    expected_values = {
        "brainvision/File_brainvision_3_int16.vhdr": {
            # SegmentDataPoints=294 in the .vhdr, a "New Segment" marker on the first sample of each segment
            "event_timestamps": {"New Segment": np.arange(34) * 294},
        },
    }

Every expected value has to come from outside neo (the specification, the vendor's reader or the file's own header) with a comment naming the source. Values copied from neo's current output would only freeze its current bugs.

I am not married to the specifics but I think we need something like this. I do something similar in neuroconv, where the EDF test class states the session start time and the subject sex in the header of edf+C.edf and the shared metadata test checks them:

https://github.com/catalystneuro/neuroconv/blob/0b1d6d4bd38f6fba982d89748bb5535731a3f921/tests/test_on_data/ecephys/test_recording_interfaces.py#L354-L373