pytest_generate_tests cannot modify fixtures of parametrized tests

In Matplotlib, we have a general autouse fixture that may or may not load a backend (i.e., UI toolkit) that causes a fatal error (if it doesn't exist), depending on test markers. In an effort to avoid that, I tried to add a fixture that skips tests that might use an unavailable backend. Because file-specific fixtures run after those in conftest.py, I used pytest_generate_tests to implement it, but this no longer seems to work on parametrized tests.

For the sake of an example, start with an autouse fixture that always fails in conftest.py:

# conftest.py
import pytest

@pytest.fixture(autouse=True)
def settings(request):
    print('settings', request.node.fixturenames)
    raise RuntimeError("This should not be called.")

then in test_foo.py, a fixture that should skip the test, inserted first via pytest_generate_tests so that it occurs before the above exception:

# test_foo.py
import pytest

def pytest_generate_tests(metafunc):
    print('pytest_generate_tests')
    metafunc.fixturenames.insert(0, 'should_skip')

@pytest.fixture
def should_skip():
    print('should_skip')
    pytest.skip('Should be skipped')

def test_foo():
    print('test_foo')
    assert True

This works for test_foo; should_skip is called first and the remaining fixtures are skipped.

If I then add a parametrized test:

@pytest.mark.parametrize('bar', [1, 2, 3])
def test_bar(bar):
    print('test_bar', bar)
    assert True

all of these call settings first instead of should_skip. The printout shows that should_skip is no longer in the list of fixtures either.

This used to work in pytest 3.6.3, but fails in 3.7.0. I bisected the issue back to #3629, but I'm not sure exactly if that intended for this behaviour to change.

</details>

Reproduce the behavior using the shown conftest.py and test_foo.py examples, comparing an unparameterized test with test_bar and the pytest 3.6.3 versus 3.7.0 behavior. Start by tracing pytest_generate_tests and fixture-name handling around the change in #3629. Done means establishing whether the regression is intentional and restoring the expected skip ordering for parametrized tests, with regression coverage.