Fix QLinearConv promotion

Currently, QDQ to QLinearConv promotion is happenning incorrectly for
per-axis quantization where axis != 0. QLinearConv can only support
per-output-channel quantization, which is equivalent to per-axis quantization
with axis=1.
The issue can be easily seen when running this model with optimization
enable (promoted to QLinearConv) and disabled (using QDQ). Both produce
different results.

onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selectors.cc

@@ -370,6 +370,13 @@ bool ConvNodeGroupSelector::Check(const GraphViewer& graph_viewer,
     return false;
   }
 
+  // Only per-tensor or per-output channel (axis == 1) quantization is supported
+  const auto& dq_attrs = dq_nodes[1]->GetAttributes();
+  if (const auto a_iter = dq_attrs.find("axis");
+      a_iter == dq_attrs.end() || a_iter->second.i() != 1) {
+    return false;
+  }
+
   return true;
 }