d3dx9_36: Use cross-platform versions of infinity and NaN.

[Dylan Smith]

On Tue, Jul 5, 2011 at 8:27 AM, Marvin <testbot at testbot.winehq.org> wrote:
> === WXPPROSP3 (32 bit math) ===
> math.c:2275: Test failed: Got ffff, expected 7fff or 7fff for index 18.

0.0f/0.0f = -nan (0xffc00000)
-(0.0f/0.0f) = nan (0x7fc00000)

msvc complains about dividing by zero. I found the following works
around the error for MS Visual Studio 2010, although later version
might try to be "smarter".

static const float z = 0.0f;
const float neg_nan = 0.0f/z;
const float pos_nan = -(0.0f/z);

Note that neg_nan and pos_nan can't be static.