MAINT: Standardize tests

[larsoner]

While working on mne-tools/mne-qt-browser#105 I kept getting unexpected errors on CIs. The first time I ran the tests locally on my macOS machine, I could replicate the issue. The second time, however, I could not. I realized this was because the MNE_BROWSE_RAW_SIZE config var on my system was being overwritten (!). This PR fixes this by:

1. Setting the _MNE_CONFIG_FAKE_HOME_DIR at the session level to prevent overwriting people's config vars
2. Setting the MNE_BROWSE_RAW_SIZE to standardize browser sizes across all systems. This should make things more repeatable in testing across different people's systems (which could otherwise have different MNE_BROWSE_RAW_SIZE). In an ideal world, this value would not matter, but in practice our _fake_clicks are not so well generalized to arbitrary window sizes.
3. Monkey-patching the native to-fullscreen methods for the two backends so that fullscreen mode is never called. This helps with testing on macOS a bit, where the fullscreen mode messes with focus a bit. (I think we can trust mpl and Qt to test their own fullscreen methods.)
4. Removing fullscreen keypress tests from epochs and ica, leaving for raw. Again I think we can trust our own abstraction here so we don't have to monkeypatch in three places.

[agramfort]

@larsoner I let you merge and backport 🙏