Fix testing of half-precision fma. (#1882)

Half-precision functions are generally tested against the
single-precision reference. This causes double rounding: first to single
precision, then from there to half precision. For the most part, it is
good enough, but specifically in the case of fma, a correctly rounded
result is required and is not obtained, for instance for arguments
0x1.eacp+7, 0x1.3f4p+4, 0x1.c04p+14, which produce an exact result of
0x1.065fffp+15 which should be rounded to half-prefcision 0x1.064p+15,
but was previously first rounded to single-precision 0x1.066p+15, and
from there to half-precision 0x1.068p+15. Testing against reference_fmal
gives us sufficient precision that double rounding does not cause
issues.

The f_fma(..., FLUSHED) calls for FTZ testing cannot be updated the same
way but do not need to be: these calls all have at least one constant
operand of zero. If one operand is zero, double rounding cannot be an
issue.

test_conformance/math_brute_force/utility.h

@@ -75,6 +75,7 @@ extern RoundingMode gFloatToHalfRoundingMode;
 extern cl_half_rounding_mode gHalfRoundingMode;
 
 #define HFF(num) cl_half_from_float(num, gHalfRoundingMode)
+#define HFD(num) cl_half_from_double(num, gHalfRoundingMode)
 #define HTF(num) cl_half_to_float(num)
 
 #define LOWER_IS_BETTER 0