[dGPU] avoid strided_slice to be executed in cpu (#25601)

### Details:
- Big input tensor to StrideSlice primitive is executed in CPU will lead
to huge performance drop.
 - *...*

### Tickets:
 - CVS-147088

---------

Co-authored-by: Pavel Durandin <pavel.durandin@intel.com>

src/plugins/intel_gpu/src/graph/graph_optimizer/mark_shape_of_subgraphs.cpp

@@ -7,6 +7,7 @@
 #include "reshape_inst.h"
 #include "eltwise_inst.h"
 #include "select_inst.h"
+#include "strided_slice_inst.h"
 #include "gather_inst.h"
 #include "pass_manager.h"
 
@@ -78,6 +79,13 @@ bool mark_shape_of_subgraphs::can_mark_node(const program_node& node) {
             return false;
     }
 
+    // Exclude stride_slice primitive if it's input is big const ternsor, else CPU reference implementation
+    // will lead to huge performance drop.
+    if (node.is_type<strided_slice>() && node.get_dependency(0).is_constant() &&
+        node.get_dependency(0).get_output_layout().count() > 1024 * 1024) {
+        return false;
+    }
+
     auto available_impls = node.type()->get_available_impls(node);
     auto cpu_impl_found = available_impls.find(impl_types::cpu) != available_impls.end();