[TRANSFORMATIONS][GPU] Add GroupNormalization fusion to common optimizations

[jhajducz]

Details:

• Added GroupNormalization fusion pass that can handle pattern observed in many customer models that were exported via ONNX in a way that uses InstanceNormalization as a proxy for GroupNormalization. It covers also more traditional cases without additional instance norm related parameters.
• Per suggestion from @vladimir-paramuzov, for now enabled GroupNormalization fusion only for GPU plugin. Once it will be verified that it doesn't cause regressions in other backends, we can enable it for them as well.

Tickets:

• 160436

[mlukasze]

build_jenkins

[t-jankowski]

Lgtm regarding Core part.

[praasz]

build_jenkins

[vladimir-paramuzov]

I think this switch can be removed and replaced with single if:

 if (!pre_mvn_shape_vals_correct(pre_mvn_shape_const->cast_vector<int64_t>(), input_ps, num_groups))

[vladimir-paramuzov]

Same here

[vladimir-paramuzov]

nit: IMO, comment here and above are not needed as the code is more or less simple

[vladimir-paramuzov]

Suggested change

	op::v0::Constant::create(element::i64, Shape{static_cast<size_t>(num_channels)}, gather_indices_vals);
	op::v0::Constant::create(element::i64, Shape{num_channels}, gather_indices_vals);

[vladimir-paramuzov]

IMO, test cases for integer types shall be removed as spec says that supported tensor types are floating-point for this op

[vladimir-paramuzov]

I think type template is not needed for this test as it can be just a parameter of the test. As I can see, in most of the cases you actually use element::Type object as a parameter. The only exception is const data generation, but you can use ov::test::utils::create_and_fill_tensor method to instead, thus the template arg won't be needed here too. That change will simplify tests instantiations I believe

[vladimir-paramuzov]

I wonder why these test cases work while fp8/fp4/bf16 types are not supported by GPU plugin at all and integer types are not supported by the kernel?

[vladimir-paramuzov]

Why do we need these args as tests parametes? I'd expect them to be hardcoded in the test

[vladimir-paramuzov]

What's the difference between this and base class version? The base one also runs model via template plugin

[vladimir-paramuzov]

I see some small difference for this impl vs base method, but IMO those differences should be handled separately to avoid copy-pasting of key execution logic. For instance, count of group norm ops can be checked in a separate method like this:

TEST_P(GroupNormalizationFusionSubgraphTestsF, CompareWithRefs) {
    run();
    check_some_plugin_or_test_specific_things_after_base_values_checks(); // add this method to GroupNormalizationFusionSubgraphTestsF
};