This is behind a configuration switch because direct unauthenticated
requests are at odds with the Matrix model where clients stick to their
homeservers and only homeservers talk to each other. The implementation
is also quite inefficient atm, causing a .well-known resolution
roundtrip every single time a direct media request is made.

Also: simplify the expression in FINISH_TEST.

Going through QSettings doesn't work quite well on Windows; setting it
in Quotest and immediately testing through another QSettings instance
inside NetworkAccessManager (in another thread) returns the previous
value. In fact, there's no general guarantee for that, as almost all
QSettings member functions are reentrant but not thread-safe. The flag
is only initialised from QSettings once at the first NAM creation, and its
changes are only propagated to QSettings but never read again.
Implication: changes in QSettings files when the client is running will
only be reflected at the next run (or may be overwritten altogether),
so just don't change settings files while the client is running, ok?

std::atomic_flag(bool) is not a standard constructor, and MSVC doesn't
have it, for one.