[NPU] Model serialization/deserialization without weights copies

[razvanapetroaie]

Details:

• Extends the OV StreamSerializer & XmlSerializer in order to allow passing an ov::Model through the driver without copying its weights into a separate buffer.
• The purpose of this PR is to reduce memory consumption by avoiding weights duplications.
• This feature will be disabled by default for the CiD interface (at least for a while). The changes are first meant to be integrated in the upcoming CiP interface.
• The implementation followed and adapted the sample provided in this PR.
• Two config options introduced: intel_npu::use_base_model_serializer - a switch between the old & new serialization algorithms, and intel_npu::serialization_weights_size_threshold - controls which weights are copied into a separate buffer and which ones have only metadata (memory location & size) stored as runtime information. More concretely, weights smaller than this value will be copied.
• vcl_serializer.hpp is meant to contain all operations required to prepare an ov::Model as food for the VCL interface. This implies using the new/old model serializer, I/O & config serialization. xml_serializer.hpp is a more generic, weightless (no weights copies unless serialization_weights_size_threshold is used) implementation of the OV serializer.
• Roughly how this works:
  • The plugin passes through all ov::Constant nodes and places weights metadata (intel_npu::WeightsPointerAttribute) as runtime information on the nodes that have buffers smaller than serialization_weights_size_threshold.
  • The new intel_npu::StreamSerialize is called which uses the intel_npu::XmlSerializer for serializing the model. Note that StreamSerialize uses a slightly different format within the buffer (metadata containing offsets & sizes, custom data, weights & the XML graph), see ov::pass::StreamSerialize for details.
  • intel_npu::XmlSerializer will not write weights into its dedicated buffer if the WeightsPointerAttribute is found within the current ov::Constant node. Instead, weights metadata will be written as runtime information by calling the visit method corresponding to the attribute.
  • The deserializer will be able to distinguish between the two cases (weights copied vs. weights stored as metadata) by looking for this attribute in the serialized buffer.
• See the ticket for some performance reports.
• TODO: some functional tests

Related PRs

• Sample for extending the serialization algorithm
• The PR that made the OV serializer extensible

Tickets:

• CVS-173711

[razvanapetroaie]

Will tune this later to see which value yields the best performance. For now, we assume 0 is the best candidate (only weights pointers & sizes are stored).

[PatrikStepan]

Decided to merge this on the master branch after the branch for OV25.4 is created. This PR alone bring no benefits to the release without further changes in the driver or without further plugin changes planned for OV26.0.