Fix GatherND broadcasting for batch dimensions

[evkotov]

Details:

Problem:
OpenVINO rejects ONNX models with GatherND when batch dimensions are broadcastable but not equal (e.g., data=[1,3,4,4], indices=[2,2,2], batch_dims=1). However, ONNXRuntime accepts and successfully executes these models.

Root Cause:

• Shape inference used strict DimType::merge() requiring exact equality of batch dimensions
• Reference implementation and CPU plugin did not support broadcasting on batch dimensions

Solution:
Implemented full broadcasting support for GatherND batch dimensions to match ONNX specification and ONNXRuntime behavior:

1. Shape Inference (gather_nd_shape_inference.hpp):

   • Replaced DimType::merge() with DimType::broadcast_merge() to allow dimension 1 to broadcast to any size N
   • Updated validation message: "must be the same" → "must be broadcastable"

2. Reference Implementation (gather_nd.hpp):

   • Compute broadcast batch shape as max(data_dim, indices_dim) for each batch dimension
   • Calculate broadcast-aware strides (stride=0 for dimension 1)
   • Multi-dimensional batch index conversion with broadcasting logic

3. CPU Plugin (gather_nd.cpp, gather_nd.h):

   • Extended GatherNDAttributes with idxDims field
   • Added broadcast support fields: batchShape, srcBatchStrides, idxBatchStrides
   • Updated both gatherBlocks() and gatherElementwise() methods with broadcast-aware iteration
   • Lambda helper getBatchOffsets() for clean broadcast offset calculation

4. Documentation (gather-nd-5.rst, gather-nd-8.rst):

   • Updated batch_dims description: "must be equal" → "must be broadcastable"

5. Tests:

   • Type prop tests: 6 new broadcasting tests (3 for v5, 3 for v8), updated fail tests
   • Shape inference tests: Added broadcasting validation test
   • Functional tests: 5 new reference tests with actual data validation (2 for v5, 3 for v8)

Tickets:

• 169059

[t-jankowski]

Please add a test in src/core/tests/type_prop/gather_nd.cpp.

[evkotov]

Please add a test in src/core/tests/type_prop/gather_nd.cpp.

I added

[mmikolajcz]

GatherND op specification mentions that batch needs to be equal, so it would also need to be updated.

Do we know whether reference/plugin implementations would handle such broadcast? I think it would make sense to add few tests to make sure result of such broadcasted GatherND would produce expected outputs.

[evkotov]

I fixed refernece implementation, docs and CPU plugin

[praasz]

Usually such change is request of new version of operator.
It has to be check if three is no compatibility with previous version of OV.
If this is for ONNX then mayb ONNX FE can process node on front end to provide compatible graph for OV

[maxnick]

@nshchego, could you please review the CPU part?

[praasz]

Usually such change is request of new version of operator.
It has to be check if three is no compatibility with previous version of OV.
If this is for ONNX then mayb ONNX FE can process node on front end to provide compatible graph for OV