Generalise Token Identity run_and_compute() method

[shaheenahmedc]

The TokenIdentity class in obvs/lenses.py currently has a run_and_compute() method, which uses batched versions of surprisal and precision. It also calculates these metrics alongside running the patchscope, saving GPU memory.

The PatchscopeLogitLens and ClassicLogitLens don't use such a method, but would benefit from it.

Can we generalise this run_and_compute() method to work for any lens? Or should it work for any patchscope?

This arose from #48.