Extend Einsum Core and common transformation to support broadcasting, repeated labels and ellipsis

[mmikolajcz]

Details:

• Extend core implementation to support extended broadcasting to comply with PyTorch op
• Extend common decomposition to support broadcasting of same label, ellipsis and repeated labels
• Repeated labels subgraph that doesn't depend on static shapes,
• Support of ellipsis label that doesn't mark any dimension,
• *Test cases"

Tickets:

• CVS-157225
• CVS-159205
• CVS-159206

[mitruska]

There are three model tests enabled in this PR, making it valuable work.
However, the complexity of the changes makes the review process non-trivial.
Please share more context about the problem that has been resolved, the solution applied to the decomposition, and the expected subgraph.

[mitruska]

Suggested change

	std::vector<std::string> common_labels, sep_labels1, sep_labels2, reduced_labels; // +++++
	std::vector<std::string> common_labels, sep_labels1, sep_labels2, reduced_labels;

[mitruska]

To call get_partial_shape().rank().get_length() static rank must be ensured, if it is ensured at this point, then get_partial_shape().size() could be used instead. Also it may be better to add to the input rank instead of subtract from the labels size, just to avoid overflow for empty labels vector.

Suggested change

	!has_ellipsis_in_input || (input_nodes[i].get_partial_shape().rank().get_length() ==
	static_cast<ov::Dimension::value_type>(labels.size() - 1));
	!has_ellipsis_in_input || ((input_nodes[i].get_partial_shape().size() + 1) == labels.size());