CUDA: fix compilation with GGML_CUDA_F16 (#14837)

ggml/src/ggml-cuda/convert.cu

@@ -31,8 +31,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = v.x;
-    y[iy0 + y_offset] = v.y;
+    y[iy0 + 0]        = float(v.x);
+    y[iy0 + y_offset] = float(v.y);
 }
 
 template <bool need_check>