LogClassifier picks error from a flaky test instead of an actual one

[malfet]

See https://hud.pytorch.org/pytorch/pytorch/commit/0478fee42db16a0477add1d0a644ce713f31a875

The actual test that caused failures in all (may be except for rocm one) are doctest failures, but classifier preferred to focus on
flaky test_python_dispatch failure