Describe the bug

In the scikit-learn test suite we have a test that launches a subprocess and the atexit registry triggers the saving of a coverage file in a tempfolder in d:\ while the current working directory has been changed by pytest --pyargs to c:\ in the meantime.

E                   AssertionError: Error in atexit._run_exitfuncs:
E                   Traceback (most recent call last):
E                     File "c:\miniconda\envs\testvenv\lib\ntpath.py", line 562, in relpath
E                       path_drive, start_drive))
E                   ValueError: path is on mount 'D:', start on mount 'c:'

cmd        = ['c:\\miniconda\\envs\\testvenv\\python.exe', 'C:\\Users\\VSSADM~1\\AppData\\Local\\Temp\\tmp32ajt78f_src_test_sklearn.py']
coverage_rc = None
cwd        = 'c:\\miniconda\\envs\\testvenv\\lib\\site-packages'
...

I do not understand why the traceback starts at atexit._run_exitfuncs and then goes to ntpath.relpath directly without going through the function of the coverage module. I suspect that the culprit is:

https://github.com/nedbat/coveragepy/blob/coverage-5.0/coverage/sqldata.py#L982

When rolling back to coverage 4.5.3, the problem disappears.

To Reproduce

This problems happens with coverage 5.0 and Python 3.7.5 with the latest pytest and pytest-cov.

It seems tricky to reproduce, in particular I do not understand why this only happens in the subprocess tests of scikit-learn and never in the main process itself. Maybe the current working directory is restored to the d:\ tempfolder prior to writing the coverage files?

Anyhow the details can be found in the CI logs of this pull request where I have patched the ntpath.relpath function to display more debug info:

scikit-learn/scikit-learn#15909

I will open a PR to fix the issue in coveragepy.

It looks similar to #824 but the error message is actually different.