Add GNN path selector

[emnigma]

This PR implements path selection with Graph Neural Network (GNN). This approach allows GNN to learn not only application context features, but also application graph structure to better predict desirable states to propagate.

GNN accepts BlockGraph with custom features as an input, uses onnxruntime framework as an inference provider. BlockGraph consists of two types of vertices: StateFeatures and BlockFeatures. They serve as an intermediate graph representation before conversion to onnx tensors.

BlockGraph is created with following rules:

• consecutive statements are put in one block
• block can't contain two external function calls, each external call is interpreted as an end of current block
• first statements after control flow branching statement start a new block
• if statement has multiple input edges it is allocated to new block

[sergeypospelov]

Please, contact @sergeyrid and unify with #60.

Main problems:

• usvm-core, usvm-jvm shouldn't depend on ML path selection. All ML-based path selectors should be located in a separate module.
• No hardcoded strings like "/Users/emax/Data/usvm/Game_env/test_model.onnx"
• Add tests, please