Security fuzz address sanitizer fix Bug #2 and microsoft#3 (microsoft#21528)

### Description
Security fuzz test with address sanitizer found several bugs

Author: jingyanwangms

onnxruntime/contrib_ops/cpu/transformers/subgraph_gpt.cc

@@ -143,6 +143,8 @@ Status GptSubgraph::Validate(const std::vector<const NodeArg*>& subgraph_inputs,
 
   // Past state shape is like (2, batch_size, num_heads, past_seq_len, hidden_size/num_heads).
   const ONNX_NAMESPACE::TensorShapeProto* past_shape = subgraph_inputs[3]->Shape();
+  ORT_RETURN_IF(past_shape == nullptr,
+                "subgraph past state cannot be nullptr");
   ORT_RETURN_IF(past_shape->dim_size() != 5,
                 "subgraph past state is expected to have 5 dimension, got ", past_shape->dim_size());
 

onnxruntime/core/optimizer/attention_fusion.cc

@@ -210,7 +210,7 @@ Status AttentionFusion::ApplyImpl(Graph& graph, bool& modified, int graph_level,
 
     if ((node.GetOutputEdgesCount() >= 2 && node.GetOutputEdgesCount() <= 6) &&  // Add node.GetOutputEdgesCount() == 5/6 for distilbert
         graph_utils::IsSupportedOptypeVersionAndDomain(node, "LayerNormalization", {1, 17}, kOnnxDomain) &&
-        graph_utils::IsSupportedProvider(node, GetCompatibleExecutionProviders())) {
+        graph_utils::IsSupportedProvider(node, GetCompatibleExecutionProviders()) && node.InputDefs().size() > 2) {
       // Get hidden size from layer norm bias tensor shape.
       const NodeArg& layer_norm_bias = *(node.InputDefs()[2]);
       if (!optimizer_utils::IsShapeKnownOnAllDims(layer_norm_bias, 1)) {