Fix concurrency issue with iterator invalidation

Author: smirnov-alexey

src/plugins/intel_npu/src/plugin/npuw/base_sync_infer_request.hpp

@@ -104,6 +104,11 @@ class IBaseInferRequest : public ov::ISyncInferRequest {
     mutable std::map<ov::Output<const ov::Node>, TensorStorage>
         m_port_to_tensor;  // mutable due to lazy I/O allocation in get_tensor()
 
+    // Have to reserve size of all structures modified by get_tensor(),
+    // otherwise iterators get invalid during parallel loops.
+    // FIXME: too much details we have to keep in mind - consider a better solution
+    void reserve_for_lazy_io();
+
     // Check that m_port_to_tensor does have a tensor stored at the port
     bool is_stored(const ov::Output<const ov::Node>& port) const;
     // Check the port is I/O

src/plugins/intel_npu/src/plugin/npuw/just_sync_infer_request.cpp

@@ -326,6 +326,7 @@ ov::npuw::JustInferRequest::JustInferRequest(const std::shared_ptr<ov::npuw::Com
         }
     }  // if(function_pipelining)
 
+    reserve_for_lazy_io();
     alloc_quant_gather();
     connect_subrequests();
     init_gio();

src/plugins/intel_npu/src/plugin/npuw/unfold_sync_infer_request.cpp

@@ -43,6 +43,7 @@ ov::npuw::UnfoldInferRequest::UnfoldInferRequest(const std::shared_ptr<ov::npuw:
         LOG_INFO("DONE");
     }  // for(submodels)
 
+    reserve_for_lazy_io();
     alloc_quant_gather();
 
     LOG_INFO("Connecting subrequests...");