At the same time, the change to the player loop prevents a buffer
overshoot and a (perhaps unlikely) program trace where a sequence
Suspend; Stop; midiStreamOut could send MOM_DONE twice and access
memory possibly released within a MOM_DONE DCB_FUNCTION callback.

Found by Valgrind.

Found by Valgrind.

Fixes a regression caused by 1ff99231.

For print preview it will take more time to wait for the whole text to be
wrapped for EM_FORMATRANGE than to wait for just the first page.  Trying to
get the total number of pages will cause the whole test to be wrapped.
Instead all that is needed is to just determine if a page is the last page,
which can be determined by it's page end offset compared to the text length.

This avoids having to delete the device context and the bitmap selected
into it just to update the bitmap, since it was setting preview.hdc to
NULL to indicate that these bitmaps needed to be updated. This patch
also helps make WM_PAINT messages get processed faster by moving this out
of the code for handling WM_PAINT.

Previously hdc2 needed to be deleted and set to NULL to indicate that only
one page should be shown. I have avoided this using a new variable for
the preview structure.

The device context were also being deleted without deleting the bitmaps
created for them and selected into them, so this patch also gets these
bitmaps and deletes them with their device context.

Rather than making seperate calls to EM_FORMATRANGE to get the number of
pages and the character boundaries, it would be easiest to get both
together.  The time to process EM_FORMATRANGE seems like it would take a
lot longer than the time needed to cache the character page boundaries.

The existing tests were not properly testing the return value for the case
where all the text fits in the page rect, so this patch adds a couple of
simple tests for this case.

winex11: When a mask picture is active, the source offsets should be relative to mask_pict instead of src_pict.