For objects that truly might not exist, use Optional.  Some objects will
always exist eventually, and for those we have some null implementation
standins to use without making new placebo classes.

Details of the problem and thought process:
https://nedbatchelder.com/blog/202302/late_initialization_with_mypy.html

Also, move tests to test_data.py, and finish covering the code.

* Add purge_files method to CoverageData, to allow for selective removal and update
of coverage data.

* Fix assert syntax so it's not true; this code isn't reached in the test unless it fails
and then it would have failed to fail.

* Remove trailing whitespace; did not expect this would matter on a blank line.

* Add type annotations required by mypy

---------

Co-authored-by: Stephan Deibel <sdeibel@wingware.com>

If the set of arcs is empty, skip the SQL operations.  We also need to
allow setting a file tracer for an unmeasured file, to avoid the Cython
problem whose fix caused the performance issue in the first place.

TBH, I don't know why we had to prevent file tracers on unmeasured
files.  Perhaps pytest-cov has changed to avoid the behavior that caused
problems.