Adding test case for conv per channel with QDQ format (microsoft#13041)

**Description**: Adding test case for conv per channel with QDQ format

Author: jchen351

onnxruntime/test/python/quantization/test_qdq.py

@@ -236,6 +236,15 @@ def test_qdq_extra_options_2(self):
 
 
 class TestQDQFormatConv(TestQDQFormat):
+    def check_per_channel_counts(self, model_path, channel_count: int, axis: int = 0):
+        model = onnx.load(Path(model_path))
+        for initializer in model.graph.initializer:
+            dims = initializer.dims
+            # skip if initializer is not a weight
+            if len(dims) > 0:
+                self.assertGreater(len(dims), axis)
+                self.assertEqual(channel_count, dims[axis])
+
     def construct_model_conv(self, output_model_path, input_shape, weight_shape, output_shape, has_bias):
         #    (input)
         #      |
@@ -281,8 +290,9 @@ def verify_quantize_conv(self, has_bias, per_channel, is_quant_type_int8=False):
         model_fp32_path = "conv_fp32.{}.{}.onnx".format(has_bias, per_channel)
         model_int8_qdq_path = "conv_quant_qdq.{}.{}.onnx".format(has_bias, per_channel)
         model_int8_qop_path = "conv_quant_qop.{}.{}.onnx".format(has_bias, per_channel)
+        channel_count = 16
         data_reader = self.input_feeds(1, {"input": [1, 8, 33, 33]})
-        self.construct_model_conv(model_fp32_path, [1, 8, 33, 33], [16, 8, 3, 3], [1, 16, 31, 31], has_bias)
+        self.construct_model_conv(model_fp32_path, [1, 8, 33, 33], [channel_count, 8, 3, 3], [1, 16, 31, 31], has_bias)
         quantize_static(
             model_fp32_path,
             model_int8_qdq_path,
@@ -301,6 +311,8 @@ def verify_quantize_conv(self, has_bias, per_channel, is_quant_type_int8=False):
             "DequantizeLinear": 4 if has_bias else 3,
         }
         check_op_type_count(self, model_int8_qdq_path, **qdq_nodes)
+        if per_channel:
+            self.check_per_channel_counts(model_int8_qdq_path, channel_count)
         check_model_correctness(self, model_fp32_path, model_int8_qdq_path, data_reader.get_next())
 
         data_reader.rewind()