squash! ggml-cuda : add ar_add() to avoid ambiguous operator+ for half/bfloat16 in CUDA 11.8

Author: danbev

ggml/src/ggml-cuda/allreduce.cu

@@ -113,7 +113,7 @@ static __device__ inline T ar_add(T a, T b) {
     if constexpr (std::is_same_v<T, half>) {
         return __hadd(a, b);
     } else if constexpr (std::is_same_v<T, nv_bfloat16>) {
-        return __hadd(a, b);
+        return __float2bfloat16(__bfloat162float(a) + __bfloat162float(b));
     } else {
         return a + b;
     }