[GPU] Handle 3D tensors by expanding to 4D in onednn reduce

[jade-cho]

Descriptions

• The purpose of PR [GPU] Use 3D shape for onednn 1D convolution #31050 was to allow oneDNN convolution to directly handle 3D input tensors without expanding them to 4D.
• However, oneDNN reduce should be excluded from this change, as it is closely tied to cldnn implementation, which expects 3D tensors to be expanded to 4D.
• This PR resolves the issue by rollback the behavior of oneDNN reduce to its previous state before PR [GPU] Use 3D shape for onednn 1D convolution #31050.

The code and line that caused this issue

• openvino/src/plugins/intel_gpu/src/graph/impls/onednn/reduce_onednn.cpp

  Line 60 in fe702ef

  auto input_md = onednn::layout_to_memory_desc(input_layout);

Problematic graph

Checklist

• Is it a proper fix? (not a workaround)
• Did you include a test case for this fix, if necessary?
  → A test case is not required, as this change is a rollback to the original behavior.
• Did you review existing tests that can be extended to cover this scenario?
  → It doesn't need.

Tickets

• 170856

[isanghao]

lgtm