[GPU] Fix reshape runtime padding propagation when cropped axis is dropped

[andrew-k-park]

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

• Symptom: GPU produced wrong box coordinates / accuracy mismatch vs CPU on a detection model with a VariadicSplit (last axis, lengths [1,1,1,1]) followed by base-mode Reshapes that drop the cropped size-1 axis ([N,M,1] -> [N,M]).
• Root cause: In is_runtime_propagatable_padding() (last-axis crop branch), the trailing-dim matching loop exits immediately when the cropped axis is size 1, returning true even when the reshape removes that axis. With no output dim left to carry the crop's dynamic padding, the four sibling reshape outputs propagated zero padding and ended up aliased to the same base-buffer offset, so downstream consumers read interleaved channel data as if it were contiguous.
• Fix: Track the number of trailing output dims actually matched against the cropped axis (matched_trailing_dims) and reject propagation when none were matched and the reshape reduces rank — i.e. the cropped axis is genuinely dropped. Identity reshapes that preserve it ([N,M,1] -> [N,M,1]) remain eligible.

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

• src/plugins/intel_gpu/src/graph/include/reshape_inst.h

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

• ./ov_gpu_unit_tests --gtest_filter='prepare_buffer_fusing.in_place_crop_dynamic_last_axis_split_to_collapsing_reshape'

Problematic graph

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 existing tests under prepare_buffer_fusing in src/plugins/intel_gpu/tests/unit/passes/prepare_buffer_fusing_test.cpp (in-place crop / reshape padding tests)

Tickets:

• CVS-185921

[andrew-k-park]

no performance regression on daily (LLM) test