Deeplima dp train: The list of expected tasks should not be hard-coded

Currently the list of tasks the parser training model expects from the tagging model (form,upos…) is hard-coded while the tagging model can be trained with various tasks.
This list should be saved somewhere and retrieved.

This is described here: https://github.com/aymara/lima/blob/7c7b1c68f4fce280e9db1aadf69a7829aacd2e0b/deeplima/libs/tasks/graph_dp/train/train_graph_dp.cpp#L77

And in 7c7b1c68f4fce280e9db1aadf69a7829aacd2e0b

If tasks lists does not match, parsing fails with
```
terminate called after throwing an instance of 'std::logic_error'
what(): Input classes missmatch: xpos != Definite
```