MµPDF: Don't accumulate rounding errors when computing page dimensions (

#1407)

* Do *NOT* round the dimensions in getSize

This is a weird MµPDF internal @ 72dpi representation, not a final
dimension.

We use it to compute and apply our own scale factor, so if we lose
precision here, we can end up with severe rounding errors with large
scale factors.

And, for some mysterious reason, for image files, scale factors can be
really large, even when the files actually match the target screen size
exactly :?.

ffi/mupdf.lua

@@ -370,9 +370,20 @@ function page_mt.__index:getSize(draw_context)
 
     M.fz_bound_page(context(), self.page, bounds)
     M.fz_transform_rect(bounds, ctm)
-    M.fz_round_rect(bbox, bounds)
 
+    -- NOTE: fz_bound_page returns an internal representation computed @ 72dpi...
+    --       It is often superbly mysterious, even for images,
+    --       so we do *NOT* want to round it right now,
+    --       as it would introduce rounding errors much too early in the pipeline...
+    -- NOTE: ReaderZooming uses it to compute the scale factor, where accuracy matters!
+    -- NOTE: This is also used in conjunction with getUsedBBox,
+    --       which also returns precise, floating point rectangles!
+    --[[
+    M.fz_round_rect(bbox, bounds)
     return bbox[0].x1-bbox[0].x0, bbox[0].y1-bbox[0].y0
+    --]]
+
+    return bounds[0].x1 - bounds[0].x0, bounds[0].y1 - bounds[0].y0
 end
 
 --[[