Be more conservative about using blending in place of masked store.

More specifically, we do a proper masked store (rather than a load-
blend-store) unless we can determine that we're accessing a stack-allocated
"varying" variable.  This fixes a number of nefarious bugs where given
code like:

    uniform float a[21];
    foreach (i = 0 … 21)
        a[i] = 0;

We'd use a blend and in turn read past the end of a[] in the last
iteration.

Also made slight changes to inlining in aobench; this keeps compiles
to ~5s, versus ~45s without them (with this change).

Fixes issue #160.

examples/aobench/ao.ispc

@@ -82,7 +82,7 @@ static inline void vnormalize(vec &v) {
 }
 
 
-static inline void
+static void
 ray_plane_intersect(Isect &isect, Ray &ray, Plane &plane) {
     float d = -dot(plane.p, plane.n);
     float v = dot(ray.dir, plane.n);
@@ -147,7 +147,7 @@ orthoBasis(vec basis[3], vec n) {
 }
 
 
-static inline float
+static float
 ambient_occlusion(Isect &isect, Plane &plane, Sphere spheres[3], 
                   RNGState &rngstate) {
     float eps = 0.0001f;