[QNN EP] Update tensor naming logic in DebugNodeInputsOutputs (microsoft#26660)

### Description
Introduced a new environment variable
ORT_DEBUG_NODE_IO_PREPEND_EP_TO_FILE_NAME to control whether the
Execution Provider (EP) type is prepended to tensor file names.
Updated MakeTensorFileName() and related logic to include the EP type in
the format: `<EPType>_<TensorName>` when the variable is set to a
non-zero value. This change helps avoid file name conflicts and improves
clarity when analyzing outputs from multi-EP runs.

### Motivation and Context
Accuracy unit tests may run the same model multiple times using
different EPs, such as ORT CPU and QNN HTP. Previously, dumped tensor
files used identical names when tensor names matched across sessions,
causing file overwrites.

### Example
When ORT_DEBUG_NODE_IO_PREPEND_EP_TO_FILE_NAME=1, dumped tensors include
EP-specific prefixes:
```
CPU_input.tensorproto
CPU_input_token_0.tensorproto
CPU_output.tensorproto
QNN_constant.tensorproto
QNN_input.tensorproto
QNN_input_token_0.tensorproto
QNN_output.tensorproto
```

Author: yuhuchua-qti

onnxruntime/core/framework/debug_node_inputs_outputs_utils.cc

@@ -99,9 +99,23 @@ struct TensorMetadata {
   std::string producer;
   std::string consumer;
   std::string device_type;
+  std::string ep_type;
   size_t step;
 };
 
+std::string GetCleanedEpType(const Node& node) {
+  std::string ep_type = node.GetExecutionProviderType();
+
+  // Remove "ExecutionProvider" suffix from ep type to reduce length.
+  const std::string suffix_to_remove = "ExecutionProvider";
+  size_t pos = ep_type.find(suffix_to_remove);
+  if (pos != std::string::npos) {
+    ep_type.erase(pos, suffix_to_remove.length());
+  }
+
+  return ep_type;
+}
+
 bool FilterNode(const NodeDumpOptions& dump_options, const Node& node) {
   auto match_pattern =
       [](const std::string& value, const std::string& delimited_patterns) {
@@ -139,14 +153,18 @@ void DumpTensorToStdOut(const Tensor& tensor, const NodeDumpOptions& dump_option
   }
 }
 
-PathString MakeTensorFileName(const std::string& tensor_name, const NodeDumpOptions& dump_options) {
+PathString MakeTensorFileName(const TensorMetadata& tensor_metadata, const NodeDumpOptions& dump_options) {
   auto make_valid_name = [](std::string name) {
     std::replace_if(
         name.begin(), name.end(), [](char c) { return !std::isalnum(c); }, '_');
     return name;
   };
 
-  return path_utils::MakePathString(make_valid_name(tensor_name), dump_options.file_suffix, ".tensorproto");
+  if (dump_options.prepend_ep_to_file_name) {
+    return path_utils::MakePathString(make_valid_name(tensor_metadata.ep_type + "_" + tensor_metadata.name), dump_options.file_suffix, ".tensorproto");
+  } else {
+    return path_utils::MakePathString(make_valid_name(tensor_metadata.name), dump_options.file_suffix, ".tensorproto");
+  }
 }
 
 void DumpTensorToFile(const Tensor& tensor, const std::string& tensor_name, const std::filesystem::path& file_path) {
@@ -375,7 +393,7 @@ void DumpCpuTensor(
       break;
     }
     case NodeDumpOptions::DataDestination::TensorProtoFiles: {
-      const std::filesystem::path tensor_file = dump_options.output_dir / MakeTensorFileName(tensor_metadata.name, dump_options);
+      const std::filesystem::path tensor_file = dump_options.output_dir / MakeTensorFileName(tensor_metadata, dump_options);
       DumpTensorToFile(tensor, tensor_metadata.name, tensor_file);
       break;
     }
@@ -485,6 +503,8 @@ const NodeDumpOptions& NodeDumpOptionsFromEnvironmentVariables() {
                 debug_node_inputs_outputs_env_vars::kHalfOverflowThreshold, " shall be a positive integer <= ", kMaxHalfThreshold);
     opts.half_overflow_threshold = static_cast<float>(threshold);
 
+    opts.prepend_ep_to_file_name = ParseEnvironmentVariableWithDefault<bool>(env_vars::kPrependEpToFileName, false);
+
     if (ParseEnvironmentVariableWithDefault<bool>(env_vars::kAppendRankToFileName, false)) {
       std::string rank = Env::Default().GetEnvironmentVar("OMPI_COMM_WORLD_RANK");
       if (rank.empty()) {
@@ -582,6 +602,7 @@ void DumpNodeInputs(
               tensor_metadata.name = input_defs[i]->Name();
               tensor_metadata.step = dump_context.iteration;
               tensor_metadata.consumer = node.Name() + ":" + std::to_string(i);
+              tensor_metadata.ep_type = GetCleanedEpType(node);
 
               TensorStatisticsData tensor_statistics;
               DumpTensor(dump_options, *tensor, tensor_metadata, tensor_statistics, session_state);
@@ -678,6 +699,7 @@ void DumpNodeOutputs(
               tensor_metadata.name = output_defs[i]->Name();
               tensor_metadata.step = dump_context.iteration;
               tensor_metadata.producer = node.Name() + ":" + std::to_string(i);
+              tensor_metadata.ep_type = GetCleanedEpType(node);
 
               TensorStatisticsData tensor_statistics;
               DumpTensor(dump_options, *tensor, tensor_metadata, tensor_statistics, session_state);

onnxruntime/core/framework/debug_node_inputs_outputs_utils.h

@@ -54,6 +54,8 @@ constexpr const char* kOpTypeFilter = "ORT_DEBUG_NODE_IO_OP_TYPE_FILTER";
 constexpr const char* kDumpDataDestination = "ORT_DEBUG_NODE_IO_DUMP_DATA_DESTINATION";
 // set to non-zero to append OpenMPI world rank to filename
 constexpr const char* kAppendRankToFileName = "ORT_DEBUG_NODE_IO_APPEND_RANK_TO_FILE_NAME";
+// set to non-zero to prepend ep type to filename
+constexpr const char* kPrependEpToFileName = "ORT_DEBUG_NODE_IO_PREPEND_EP_TO_FILE_NAME";
 // specify the output directory for any data files produced
 constexpr const char* kOutputDir = "ORT_DEBUG_NODE_IO_OUTPUT_DIR";
 // specify the file prefix for sqlite3 db (process id will be appended)
@@ -117,6 +119,9 @@ struct NodeDumpOptions {
     SqliteDb
   } data_destination{DataDestination::StdOut};
 
+  // Whether to prepend ep type to output file name.
+  bool prepend_ep_to_file_name{false};
+
   std::string file_suffix;
   // the output directory for dumped data files
   std::filesystem::path output_dir;