The tests were overly restrictive.  The functions are supposed to
return values less than 0, 0, or greater than 0, whereas the tests
restricted the return values to -1, 0, or 1.  Using less than
0/greater than 0 comparisons rather than == -1 or == 1 should fix a
few failing tests, and match the API descriptions better.

gdiplus: In GdipDrawString do not NULL terminate the buffer. It is unneeded and overruns the buffer.

This is as much an experiment as anything.  The tests currently time
out, so I'm curious whether fixing some other failures will result in
the timeout going away, or at least moving elsewhere.

kernel32: Log more information in case of a test failure, and don't run tests that are guaranteed to fail.

setupapi: Add stub entries for CM_Query_And_Remove_SubTreeA/W and CM_Query_And_Remove_SubTree_ExA/W.