disable broadcast on flash_attn_ext

Author: chraac

ggml/src/ggml-qnn/npu/device/op_flash_attn.cpp

@@ -312,6 +312,14 @@ bool is_flash_attn_supported(npu_device_tensor_op op, const npu_device_tensor_sp
         return false;
     }
 
+    if (q->ne[2] != k->ne[2] || q->ne[3] != k->ne[3] || q->ne[3] != 1) {
+        // TODO: add broadcast support
+        DEVICE_LOG_DEBUG("[%s]q and k shapes do not match: q ne: %ld, %ld, %ld, %ld, k ne: %ld, %ld, %ld, %ld\n",
+                         op_get_name(op), q->ne[0], q->ne[1], q->ne[2], q->ne[3], k->ne[0], k->ne[1], k->ne[2],
+                         k->ne[3]);
+        return false;
+    }
+
     return true;
 }