RandR matrix computations lose too much precision in fixed point;
computations using the inverted matrix can be as much as 10 pixels off.
Convert them to double precision values and pass those around. These API
changes are fairly heavyweight; the official Render interface remains fixed
point, so the fixed point matrix comes along for the ride everywhere.
This reduces the matrix representation error after inverting a
transformation matrix (although it doesn't eliminate it entirely).
Perhaps we should extend Render to include 64-bit floating point transforms...
The render extension uses many matrix operations internally, this change
exposes those functions to other parts of the server, drivers and
extensions. The change is motivated by the 'transform' additions to the
RandR extension but will likely be useful elsewhere.