Improper toolkit resolution variant scale should be properly handled

See the discussion: http://mail.openjdk.java.net/pipermail/2d-dev/2016-September/007660.html
-------------
But we still have the issue that the *actual* image size may not be
what they said it was - either explicitly or by convention.

Supposing what is claimed to be a 1.5x1.5 scale image is actually
1.0x2.0 times the size of the base image ? It is not even uniform.

Ultimately what needs to "win" is the w:h ratio of the base image
and we generally would want to pick whichever image best works
for the actual device scale, based on the *real* dimensions of
the hi-res image, don't we ?
-------------