[NPU] Model marshalling: fallback to copying weights for all unhandled Constant nodes (#33068)

### Details:
- The [new model marshalling
algorithm](#31939) is
not handling subgraphs properly. See [this code
passage](https://github.com/openvinotoolkit/openvino/blob/0f37471fb697a07c46f03c01dedb624c85628b34/src/plugins/intel_npu/src/compiler_adapter/src/vcl_serializer.cpp#L152-L163):
the attribute is added only on `ov::Constant` nodes found at depth level
one. Models within models, which may be registered as "subgraph
operators" (e.g. `TensorIterator`) are omitted. In this current state,
the attribute is required on all constant nodes, otherwise the run will
fail at the deserialization stage.
- As a temporary fix to unblock [this
PR](#30732), all
constant nodes that have been omitted will have their weights copied in
a separate buffer. We may consider this a safety measure for now, with
no real impact on performance. The impact can be great if, say, a model
has a huge submodel within it, but the data we've gathered so far
suggests these cases are not frequent (none of our models from the CI
validation jobs use this feature).

### Tickets:
 - *CVS-177414*

Author: razvanapetroaie

src/plugins/intel_npu/src/compiler_adapter/include/xml_serializer.hpp

@@ -40,13 +40,23 @@ class XmlSerializer : public ov::util::XmlSerializer {
                                   false,
                                   ov::element::dynamic,
                                   false),
+          m_base_constant_writer(std::ref(constant_write_handler)),
           m_weightless_constant_writer(weightless_constant_writer
                                            ? weightless_constant_writer
                                            : std::make_shared<WeightlessWriter>(constant_write_handler)) {}
 
 private:
+    /**
+     * @brief Toggles between the two writers.
+     */
     ov::util::ConstantWriter& get_constant_write_handler() override;
 
+    /**
+     * @brief Overriden in order to choose which weights writer will be used based on the occurrence of the
+     * "WeightsPointerAttribute".
+     */
+    bool append_node_attributes(ov::Node& node) override;
+
     std::unique_ptr<ov::util::XmlSerializer> make_visitor(pugi::xml_node& data,
                                                           const std::string& node_type_name,
                                                           ov::util::ConstantWriter& constant_write_handler,
@@ -56,7 +66,15 @@ class XmlSerializer : public ov::util::XmlSerializer {
                                                           ov::element::Type,
                                                           bool) const override;
 
+    /**
+     * @brief The base OV writer, copies the weights in a dedicated buffer.
+     *
+     * @note Ideally, we would not require this writer at all. The current algorithm does not handle subgraphs properly,
+     * so falling back to copying a part of the weights is a temporary fix.
+     */
+    std::reference_wrapper<ov::util::ConstantWriter> m_base_constant_writer;
     std::shared_ptr<WeightlessWriter> m_weightless_constant_writer = nullptr;
+    bool m_use_weightless_writer = false;
 };
 
 /**

src/plugins/intel_npu/src/compiler_adapter/src/xml_serializer.cpp

@@ -10,7 +10,20 @@
 namespace intel_npu {
 
 ov::util::ConstantWriter& XmlSerializer::get_constant_write_handler() {
-    return *m_weightless_constant_writer;
+    if (m_use_weightless_writer) {
+        return *m_weightless_constant_writer;
+    } else {
+        return m_base_constant_writer;
+    }
+}
+
+bool XmlSerializer::append_node_attributes(ov::Node& node) {
+    // If the "WeightsPointerAttribute" is found, then we have the metadata required to avoid copying the weights
+    // corresponding to this node.
+    m_use_weightless_writer = node.get_rt_info().count(WeightsPointerAttribute::get_type_info_static()) != 0;
+    auto result = ov::util::XmlSerializer::append_node_attributes(node);
+    m_use_weightless_writer = false;
+    return result;
 }
 
 std::unique_ptr<ov::util::XmlSerializer> XmlSerializer::make_visitor(pugi::xml_node& data,