[GPU] fix performance drop for crop in yolo11 model

[yuanxion]

Details

• Fix performance regression introduced by PR [GPU] Fix accuracy degradation issue for hbonet0.5/hbonet1.0/nanodet-m-1.5x-416 models caused by eltwise kernels #31140.

Description of the issue

Symptom

manual_yolo11 model performance dropped from 362.4 FPS to 318.25 FPS on GPU.

Root cause

• previous PR will force all crop primitives followed by onednn concatenation to clean its GPU memory by filling with zeros if it is blocked format.
• manual_yolo11 model also has many such crop primitives, so its performance will drop.

How to fix it

• Found that filling GPU memory with zeros can be skipped if the crop primitive uses eltwise_blocked_opt kernel and is not dynamic, so just skip it by checking crop primitive's kernel name.

The code and line that caused this issue

openvino/src/plugins/intel_gpu/src/graph/primitive_inst.cpp

Line 2036 in 453c8ee

if (get_node().is_type<crop>() && get_node().can_share_buffer() && _impl_params->get_output_layout(0).format.is_blocked() &&

Reproduction step and snapshot

• benchmark_app
  benchmark_app -inference_only false -b 1 -t 60 -nireq 4 -d GPU.0 -hint none -nstreams 2 -m INT8/1/ov/optimized/manual_yolo11.xml

Problematic graph

• crop primitive (eltwise_blocked_opt kernel) followed by onednn concatenation in manual_yolo11

Checklist

• Is it a proper fix? (not a workaround)
• Did you include test case for this fix, if necessary? No need
• Did you review existing test that can be extended to cover this scenario? Which test did you review?

Tickets:

• CVS-173402

[yuanxion]

Also checked that the change in this PR will not cause accuracy drop for hbonat/nanodet models in PR #31140.