Quantize Weight for Gemm/Conv on Quantized Model (#22969)

Some quantized models have QDQ around Conv/Gemm but the weight and/or
bias are not quantized. This PR adds WeightBiasQuantization optimizer to
quantize float weight and/or bias to INT8 and INT32 tensors
respectively. We only do this for weight and/or bias initializer so that
ConstantFolding will fold the sub-graph to real quantized initializers
during the graph optimization next round.

Author: Lafi7e

onnxruntime/core/optimizer/graph_transformer_utils.cc

@@ -63,7 +63,7 @@
 #ifdef MLAS_TARGET_AMD64_IX86
 #include "core/optimizer/qdq_transformer/avx2_weight_s8_to_u8.h"
 #endif
-#include "core/optimizer/qdq_transformer/bias_quantization.h"
+#include "core/optimizer/qdq_transformer/weight_bias_quantization.h"
 #include "core/optimizer/qdq_transformer/clip_quantizelinear.h"
 #include "core/optimizer/qdq_transformer/ensure_unique_dq_for_node_unit.h"
 #include "core/optimizer/qdq_transformer/qdq_propagation.h"
@@ -245,7 +245,7 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
 
       if (!disable_quant_qdq) {
         transformers.emplace_back(std::make_unique<QDQPropagationTransformer>());
-        transformers.emplace_back(std::make_unique<BiasQuantization>());
+        transformers.emplace_back(std::make_unique<WeightBiasQuantization>());
 
         // EnsureUniqueDQForNodeUnit is actually a required graph transformation. The unique DQ per QDQ node unit input
         // condition that it ensures is important for the partitioning that happens after Level1 optimizers are run.