fix(core): gf16_from_f32(NaN) returns NaN instead of +inf (#45)

fromF32 collapsed NaN into infinity because the non-finite branch
set mant=0 (the infinity encoding) without checking isNan first.
Split into separate isNan and isInf checks; NaN now encodes as
exp=0x3F, mant=1 (canonical GF16_NAN = 0x7E01).

Add C-ABI test for NaN preservation round-trip.

Closes #38

Author: gHashTag

src/c_abi.zig

@@ -285,6 +285,10 @@ test "C-ABI: gf16_is_nan and gf16_is_inf" {
 
     const zero = gf16_from_f32(0.0);
     try std.testing.expect(gf16_is_zero(zero));
+
+    const nan_val = gf16_from_f32(std.math.nan(f32));
+    try std.testing.expect(gf16_is_nan(nan_val));
+    try std.testing.expect(!gf16_is_inf(nan_val));
 }
 
 test "C-ABI: gf16_phi_quantize" {

src/formats/golden_float16.zig

@@ -88,7 +88,11 @@ pub const GF16 = packed struct(u16) {
     pub fn fromF32(v: f32) GF16 {
         if (v == 0.0) return .{ .mant = 0, .exp = 0, .sign = 0 };
 
-        if (!std.math.isFinite(v)) {
+        if (std.math.isNan(v)) {
+            return .{ .mant = 1, .exp = 0x3F, .sign = 0 };
+        }
+
+        if (std.math.isInf(v)) {
             return .{ .mant = 0, .exp = 0x3F, .sign = @intFromBool(v < 0) };
         }