[CPU][DEBUG_CAPS] Bring back an execution of all nodes for debug caps (…

…#28597)

The logic was reverted due to the issues with making all the nodes
executable in some scenarious.
The problem is that without this functionality debug capabilities became
less usable.
So bring back the logic and try to cover problematic scenarios,
i.e. static nodes with empty shapes, which should never be executed
even with debug caps.

src/plugins/intel_cpu/src/graph.cpp

@@ -288,12 +288,16 @@ static std::tuple<std::vector<NodePtr>, std::vector<size_t>> ExtractExecutableNo
     std::unordered_map<size_t, size_t> graphIdToExecutableId;
     std::vector<NodePtr> executableGraphNodes;
     for (size_t i = 0; i < graphNodes.size(); i++) {
-        const auto& graphNode = graphNodes[i];
-        if ((!graphNode->isConstant() && graphNode->isExecutable()) ||  // non-constant executable or
-            (graphNode->isDynamicNode() &&
-             !one_of(graphNode->getType(), Type::Input, Type::Output))) {  // dynamic, except inputs / outputs
+        const auto& node = graphNodes[i];
+        const bool staticZeroDims = !node->isDynamicNode() && !node->isExecutable() && !node->isInPlace();
+        const bool dynamicNonInputOutput = node->isDynamicNode() && !one_of(node->getType(), Type::Input, Type::Output);
+
+        if (!node->isConstant() &&  // constants are executed once in scope of compile_model
+            !staticZeroDims &&      // never execute static nodes with zero dim input / output tensors
+            (CPU_DEBUG_CAPS_ALWAYS_TRUE(node->isExecutable()) ||  // execute all executable nodes
+             dynamicNonInputOutput)) {                            // plus dynamic ones, except inputs / outputs
             graphIdToExecutableId[i] = executableGraphNodes.size();
-            executableGraphNodes.emplace_back(graphNode);
+            executableGraphNodes.emplace_back(node);
         }
     }
 

src/plugins/intel_cpu/src/utils/debug_capabilities.h

@@ -181,6 +181,8 @@ static inline std::ostream& _write_all_to_stream(std::ostream& os, const T& arg,
 
 #    define CREATE_DEBUG_TIMER(x) PrintableTimer x
 
+#    define CPU_DEBUG_CAPS_ALWAYS_TRUE(x) true
+
 /*
  * important debugging tools for accuracy issues
  *   OV_CPU_INFER_PRC_POS_PATTERN : positive regex pattern to filter node type & orgname.
@@ -286,6 +288,8 @@ bool getEnvBool(const char* name);
 
 #    define CREATE_DEBUG_TIMER(x)
 
+#    define CPU_DEBUG_CAPS_ALWAYS_TRUE(x) x
+
 #endif  // CPU_DEBUG_CAPS
 
 // To avoid "unused variable" warnings `when debug caps