Reduce bidi duplications from gdi32 by using the newly corrected usp10 
functions.

This is safe because the reference count is stored in 2 bytes on disk.

(This is commit ff2f8948 reapplied
because of incorrect revert in 88214a72).

While some performance hit from the extra flushes was expected and would be
acceptable, the performance hit can be over 50% on some configurations, which
clearly isn't acceptable. This patch introduces a "StrictDrawOrdering"
registry key for applications affected by this, until we have a more
structural solution.