Streamline ML benchmarks

[leggerf]

As discussed, we can unify the ML benchmarks for better maintenance.

[gabrielefronze]

Note from meeting: make all the ML TF benchmarks a single benchmark made of multiple presets. As such all the benchmarks can be based on the same base class.

[leggerf]

Also refactored the documentation