Create test_irs.py to detect changes in tiny-model IRs

[anatyrova]

What does this PR do?

This PR implements test_irs.py to detect changes in tiny-model IRs against reference IRs.
It will allow us to accelerate moving to new transformers version by default and check that IRs generated for newer transformers versions stay performant. Git LFS is implemented to store the reference IRs, that are generated with latest for optimum-intel 5.5.4 transformers.

P.S. more tiny models will be added soon here

pytest tests/openvino/test_irs.py -v

eg. small part of failed traceback:

E             Not equal output element types of __module.encoder.block.4.layer.1.layer_norm/aten::mul/Multiply_1 and Constant_35780 on 0 index model_one: <Type: 'float32'>, model_two: <Type: 'int64_t'>.
E             Not equal op partial shapes of self.encoder.block.4.layer.1.DenseReluDense.wi.weight and __module.encoder.final_layer_norm/aten::mean/ReduceMean on 0 index model_one: [37,32], model_two: [?,?,1].
E             Not equal op partial shapes of __module.encoder.block.4.layer.1.DenseReluDense.wi/aten::linear/MatMul and Constant_39730 on 0 index model_one: [?,?,37], model_two: [1,1,1].
E             Not equal op partial shapes of __module.encoder.block.4.layer.1.DenseReluDense.act/aten::relu/Relu and __module.encoder.final_layer_norm/aten::add/Add on 0 index model_one: [?,?,37], model_two: [?,?,1].
E             Not equal op partial shapes of self.encoder.block.4.layer.1.DenseReluDense.wo.weight and __module.encoder.final_layer_norm/aten::rsqrt/Sqrt on 0 index model_one: [32,37], model_two: [?,?,1].
E             Not equal op partial shapes of __module.encoder.block.4.layer.1.DenseReluDense.wo/aten::linear/MatMul and __module.encoder.final_layer_norm/aten::rsqrt/Divide on 0 index model_one: [?,?,32], model_two: [?,?,1].
E             Not equal op partial shapes of Constant_40674 and __module.encoder.final_layer_norm/aten::mul/Multiply_1 on 0 index model_one: [1,1,1], model_two: [?,?,32].
E             Not equal op partial shapes of Constant_40672 and Result_35882 on 0 index model_one: [1,1,1], model_two: [?,?,32].
E           
E           Reference IR was generated with:
E             - transformers: 5.5.4
E             - optimum-intel: 2.1.0.dev0+fd94990
E             - openvino: 2026.2.1-21919-ede283a88e3-releases/2026/2
E             - date: 2026-08-03
E           
E           This may indicate a regression or improvement in IR generation. Review the changes and update reference IRs if the change is intentional.

tests/openvino/test_irs.py:222: Failed

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[rkazants]

@anatyrova
please revise changes, something wrong +133,707

[rkazants]

please remove json files

[rkazants]

how long will it take if we generate reference IRs during testing time?