freestanding libc: export fmal

libc requires this to use `long double` which is sometimes the same as
f128, sometimes not.

Also for an unknown reason, aarch64 is getting an invalid result for the
`@mulAdd` behavior test for f128. See #9900.

lib/std/math/fma.zig

@@ -15,6 +15,10 @@ pub fn fma(comptime T: type, x: T, y: T, z: T) T {
         f32 => fma32(x, y, z),
         f64 => fma64(x, y, z),
         f128 => fma128(x, y, z),
+
+        // TODO this is not correct for some targets
+        c_longdouble => @floatCast(c_longdouble, fma128(x, y, z)),
+
         else => @compileError("fma not implemented for " ++ @typeName(T)),
     };
 }

lib/std/special/c_stage1.zig

@@ -656,6 +656,10 @@ export fn ceil(x: f64) f64 {
     return math.ceil(x);
 }
 
+export fn fmal(a: c_longdouble, b: c_longdouble, c: c_longdouble) c_longdouble {
+    return math.fma(c_longdouble, a, b, c);
+}
+
 export fn fma(a: f64, b: f64, c: f64) f64 {
     return math.fma(f64, a, b, c);
 }

test/behavior/muladd.zig

@@ -24,7 +24,8 @@ fn testMulAdd() !void {
         var c: f64 = 6.25;
         try expect(@mulAdd(f64, a, b, c) == 20);
     }
-    {
+    // TODO https://github.com/ziglang/zig/issues/9900
+    if (@import("builtin").cpu.arch != .aarch64) {
         var a: f16 = 5.5;
         var b: f128 = 2.5;
         var c: f128 = 6.25;