Add fast versions of the float<-->half conversion routines in the stdlib.

These get slightly wrong results for zero and the denorms and also
don't handle the Inf/NaN stuff correctly, but are much more efficient
than the full versions of these routines.

tests/half-2.ispc

@@ -0,0 +1,32 @@
+
+export uniform int width() { return programCount; }
+
+export void f_v(uniform float RET[]) {
+    float sum = 0;
+    int errors = 0;
+
+    // We should match up except for the denorms (0->1023) and the inf/nan
+    // ranges.
+    for (uniform int i = 1024; i < 32768-1024; ++i) {
+        unsigned int16 h = i;
+        float f = half_to_float_fast(i);
+        h = float_to_half_fast(f);
+
+        if (i != h)
+            ++errors;
+    }
+
+    for (uniform int i = 32768+1024; i < 65536-1024; ++i) {
+        unsigned int16 h = i;
+        float f = half_to_float_fast(i);
+        h = float_to_half_fast(f);
+
+        if (i != h)
+            ++errors;
+    }
+    RET[programIndex] = errors;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 0;
+}