[GPU] Fix in-place crop optimization for batch-axis split with spatial-flatten reshape

[andrew-k-park]

Description of the issue

• Symptom: RAFT Large (torchvision optical flow) produces completely wrong inference results on GPU when model.reshape() is called to set a static batch size. SSIM drops from 0.999 to 0.903, with max absolute error ~3.86 across all 12 outputs.
• Root-cause: Commit 222c9d4 ("Extend in-place crop optimization to support batch-axis split with reshape") only checked crop_axis == 0 && batch == 1 to identify batch-squeeze reshapes. This incorrectly matched RAFT's spatial-flatten reshape pattern [1, 256, 65, 120] → [1, 256, 7800] where the batch dim is preserved but spatial dims are merged. The wrong dyn_pad_dims padding was applied to the crop and its downstream reshape, causing them to read data from incorrect memory offsets.
• Resolution: Added a guard that checks the reshape output's first dimension: if it is statically 1, the batch dim was not squeezed (spatial-flatten case) and the batch-axis optimization is skipped. The condition !(reshape_output[0].is_static() && reshape_output[0].get_length() == 1) is applied at all three code paths: is_runtime_propagatable_padding(), build-time dynamic padding mask, and runtime padding offset calculation.

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

• intel_gpu/src/graph/include/reshape_inst.h
• intel_gpu/src/graph/graph_optimizer/prepare_buffer_fusing.cpp

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

• $ pytest test_ovc_mo.py --modules pipelines/production/pytorch/heavy -k PyTorch_TorchvisionRaftLarge --dynamism_type=None --log-cli-level INFO

Problematic graph

• RAFT contains the following subgraph repeated in its correlation/update blocks:
  • The optimization incorrectly treated this as the batch-squeeze pattern ([1,C,H,W] → [C,H,W]) and applied batch-axis dynamic padding, causing crop nodes to be marked optimized out with wrong dyn_pad_dims.
  • In the graph dump, exactly 2 nodes differed from the reference: VariadicSplit.out1 (crop) and Reshape_1 changed from nopad to dyn_pad_dims.

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 in_place_crop_dynamic_batch_axis_split_with_reshape — this tests the valid batch-squeeze pattern and Added new test

Tickets:

• CVS-182560