Avoid calling threadPaused() on exit from STG land if we're just
switching to the interpreter, and conversely call threadPaused() in
the interpreter if we're returing to the scheduler for anything other
than switching to STG.

This will probably fix the recent slowdown in GHCi (ioref001 test, for
example).  It was broken when we moved the threadPaused() call into
STG from the scheduler, so it only affects the HEAD.