`InThreadIMDServer` should bind a free port and return it rather than using `get_free_port`

[ljwoods2]

Expected behavior

InThreadIMDServer binds a free port and makes it available somehow to the IMDClient that attempts to connect to it.

Actual behavior

InThreadIMDServer takes a port as an argument and in testing we get a port using get_free_port(). But in the space between the call to get_free_port and instantiation of the server, another process can use the free, unbound port, creating a rare but annoying race condition that will make the test fail during setup.

This impacts the core MDA codebase now, too, since the PR MDAnalysis/mdanalysis#4923 uses the test server