Fix build with LLVM 23

Before LLVM 24, __builtin_amdgcn_fdot2_f32_bf16 takes short2 operands, and
implicit conversions between vector types are rejected.
--- a/src/include/miopen/conv/device_grouped_conv_fwd.hpp
+++ b/src/include/miopen/conv/device_grouped_conv_fwd.hpp
@@ -17,7 +17,11 @@
 {
 #if defined(__gfx950__)
     // gfx950 has a hardware bf16 dot; matches CK's half2_t path (v_dot2_f32_f16).
+#if __clang_major__ >= 24
     typedef __bf16 bf16x2_t __attribute__((ext_vector_type(2)));
+#else
+    typedef short bf16x2_t __attribute__((ext_vector_type(2)));
+#endif
     c = __builtin_amdgcn_fdot2_f32_bf16(
         __builtin_bit_cast(bf16x2_t, a), __builtin_bit_cast(bf16x2_t, b), c, false);
 #else