feat(arch): add the amd64 extended-instruction layer with BF16 and VP2INTERSECT

Assisted-by: GLM 5.3 Flash

arch/arm64_ext_test.go

@@ -386,10 +386,10 @@ func TestArm64ExtRejects(t *testing.T) {
 }
 
 // TestExtensionsArchBinding pins the registry's architecture binding: the
-// extended layer exists for arm64 alone until an amd64 table attaches, and no
-// other architecture sees a single SVE instruction.
+// extended layer exists for arm64 and amd64 (the latter in amd64_ext.go), and
+// no other architecture sees a single instruction of either.
 func TestExtensionsArchBinding(t *testing.T) {
-	for _, a := range []Arch{AMD64, RISCV, LOONG64, Unknown} {
+	for _, a := range []Arch{RISCV, LOONG64, Unknown} {
 		if got := Extensions(a); len(got) != 0 {
 			t.Errorf("Extensions(%s) carries %d instructions, want none", a, len(got))
 		}
@@ -397,4 +397,7 @@ func TestExtensionsArchBinding(t *testing.T) {
 	if got := Extensions(ARM64); len(got) == 0 {
 		t.Error("Extensions(ARM64) is empty")
 	}
+	if got := Extensions(AMD64); len(got) == 0 {
+		t.Error("Extensions(AMD64) is empty")
+	}
 }