Added 5 parameters plus validation of their combinations.

Also test case for the validation logic.

Validation is simply checking they exist.

Also test case

Secure by default

Whoa, we never asserted this before...

It is ignored anyways if --tlscert and --tlskey not specified.

This makes them easily importable by test suites.

Also create an __all__ module attribute listing all the variables (excl.
fixtures) available in conftest.py

... and their test cases.

* test_ssl_files renamed to test_ssl_files_err
* Use constants for test_ssl_files_err
* isort
* black

When run using pytest directly, 1.0 is enough.

But tox (and probably PyCharm tox runner) added enough overhead that 1.0
is barely enough, resulting in flakey test.

A bit of difference between tox & GA:

* In tox, Bandit is run in "qa", because Bandit can run on Windows and
  if placed in "static", it will not be run, and that's a loss.
* In GA, Bandig is run during "Static Code Checking" because that's more
  accurate.

Lines marked "# nosec" are intentional, so we suppress Bandit's
complaints on those lines.

pypy3.6-7.3.1 puked on this

According to Bandit docs in https://github.com/PyCQA/bandit, we should
run Bandit using the same Python version as will be used by the program.

Since aiosmtpd is meant to be a cross-platform, cross-version library,
we move it up there to Bandit-test with every supported version.

* Extract test_tls, test_tls_noreq, and test_smtps into a separate class
  to better control the event loop
* Change from "naive timeout" (which sometimes result in test-flakines
  to Event-based continuation (more deterministic)
* Schedule the autostop to as near as possible to main() call

Because the watch_for_* watchers are now much more deterministic after
we change over to Event-based-continuation.

See 2c0bc26 for reasoning

PyPy3.7.9-7.3.3 still barfs here.

Apparently Unicode triangles is not as universal as the checkerboard
pattern. Go figure.