[WebGPU EP] bug fix for convolution operator (microsoft#25000)

`is_channels_last` is being passed to MatMulProgram but not to
MatMulNaiveProgram causing issues for musicgen model

Author: prathikr

onnxruntime/core/providers/webgpu/math/matmul.cc

@@ -55,7 +55,7 @@ Status MatMulNaiveProgram::GenerateShaderCode(ShaderHelper& shader) const {
   std::string process_bias;
   if (has_bias_) {
     shader.AddInput("bias", ShaderUsage::UseUniform);
-    process_bias = is_channels_last_ ? "value += output_value_t(bias[col])" : "value += output_value_t(bias[row + i]);";
+    process_bias = is_channels_last_ ? "value += output_value_t(bias[col]);" : "value += output_value_t(bias[row + i]);";
   }
 
   std::string apply_activation = GetActivationSnippet(activation_, "output_value_t", "output_element_t");

onnxruntime/core/providers/webgpu/nn/conv.cc

@@ -192,7 +192,7 @@ Status Conv<is_channels_last, is_fused>::ComputeInternal(ComputeContext& context
       uint32_t output_size = static_cast<uint32_t>(output_shape.Size() / components / output_number);
       const size_t output_rank = matmul_output_shape.NumDimensions();
       TensorShape outer_dims = output_rank > 2 ? matmul_output_shape.Slice(0, output_rank - 2) : TensorShape({});
-      MatMulNaiveProgram program(activation_, output_rank, output_number, has_bias);
+      MatMulNaiveProgram program(activation_, output_rank, output_number, has_bias, is_channels_last);
       program
           .CacheHint(std::to_string(components), std::to_string(a_components), std::to_string(output_number))
           .AddInputs({{matmul_inputs[0], ProgramTensorMetadataDependency::TypeAndRank, ReduceShapeByComponents(matmul_input_reshapes[0], a_components), int(a_components)},