Instead of simply rounding layer origins and discarding subpixel
offsets, accumulate them by transforming them into the space of the
next child stacking context. This is an attempt to eliminate subpixel
differences that are caused by different stacking context boundaries in
reference tests.
Currently these accumulated subpixels are only used for text
positioning, but the plan is that they can be used for all drawing in
the future.
Without this change, each text fragment in a block that establishes a
stacking context will establish its own stacking context. This is
unnecessary and increases the amount of work done during display list
construction.
