Improve the precision of DirectInfluence with second order mode Gauss Newton

As discussed in #591 it might be possible to build up the inverse of the Gauss Newton matrix directly without materializing the Gauss Newton matrix (as currently done [here](https://github.com/aai-institute/pyDVL/blob/c1ef37bbb682a02f891431989180ea94653fa0aa/src/pydvl/influence/torch/influence_function_model.py#L455)).

On top of a potential speedup, this could increase the precision as it avoids Sin 2 of the [7 sins of numerical linear algebra](https://nhigham.com/2022/10/11/seven-sins-of-numerical-linear-algebra/)

The issue references a specific file: src/pydvl/influence/torch/influence_function_model.py line 455, where the Gauss Newton matrix is currently materialized. Start by understanding the existing DirectInfluence implementation and the mathematical approach for building the inverse directly. Review the linked issue #591 for context. 'Done' means implementing a more precise and potentially faster second-order mode without materializing the full matrix, and verifying correctness against existing tests.

Scritto dal modello di indicizzazione a partire dal testo della issue.