[GPU] Skip in-place concat when a predecessor is in a shape-of subgraph

[andrew-k-park]

Description of the issue (symptom, root-cause, how it was resolved)

• Symptom: [GPU] Padded output is not supported yet thrown from crop_cpu_impl during inference when a concat mixes a shape-of subgraph input with a regular one. Reproduced on GPU with a model that has a shape-of subgraph (ShapeOf -> ... -> VariadicSplit) inside a Loop body.
• Root cause: concat_in_place_optimization forces offset padding on its predecessor outputs so they can write directly into the concat output buffer. Nodes inside a shape-of subgraph run on CPU impls (e.g. crop lowered from VariadicSplit) which do not support padded output. The existing guard only checked concat_node.is_in_shape_of_subgraph() — when the concat itself is outside the subgraph but one of its predecessors is inside it, the optimization still fired and the CPU crop impl asserted at runtime.
• Resolution: In the predecessor loop of concat_in_place_optimization::match, reject the optimization when any predecessor is_in_shape_of_subgraph(). The concat then uses an explicit (non-in-place) implementation, so no padding is forced on the CPU-impl predecessor. Behavior for non-shape-of predecessors is unchanged.

The code and line that caused this issue (if it is not changed directly)

• intel_gpu/src/graph/impls/cpu/crop.cpp:65
  • OPENVINO_ASSERT(!padded_output, "[GPU] Padded output is not supported yet") — CPU crop impl cannot produce a padded output.
• intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp
  • concat_in_place_optimization::match() — existing guard checked only the concat node, not its predecessors.

Reproduction step and snapshot (if applicable. Do not attach for customer model)

Minimal graph inside a shape-of subgraph reaching the CPU crop impl:

input -> ShapeOf -> Gather -> VariadicSplit -> crop.out1[3] -+   (CPU impl, in shape-of subgraph)
                                                             |
                          eltwise (out of subgraph) ---------+-> Concat -> ...

Concat is outside the subgraph → passes the existing guard → in-place optimization forces padded output on crop.out1 → CPU crop impl throws at runtime.

Regression test: ov_gpu_unit_tests --gtest_filter='prepare_buffer_fusing.in_place_concat_rejected_for_shape_of_subgraph_input'

Problematic graph

                    shape-of subgraph (CPU impls)
       +------------------------------------------------+
input->| ShapeOf -> Gather -> VariadicSplit -> crop1 ---|--+
       +------------------------------------------------+  |
                                                           +-> Concat -> ...
                       eltwise (outside subgraph) ---------+

crop1 is a CPU-impl node inside the shape-of subgraph; Concat sits outside it. In-place concat makes crop1 write with an offset into the shared buffer (padded output), which the CPU crop impl does not support.

Checklist

• Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary?
• Did you review existing test that can be extended to cover this scenario? Which test did you review? — Reviewed prepare_buffer_fusing_test.cpp; existing in_place_concat_* tests use regular (non-shape-of) predecessors and don't exercise the CPU-impl padded-output path, so a new test was added.

Tickets:

• CVS-190378