Benchmarks updates & Added Blink benchmark support

[GuilhermeViveiros]

This PR introduces several updates:

There has been significant work within deep-spin that, unfortunately, was not merged periodically into main.
I am unable to provide a detailed description of all the changes, but I can highlight the main points.
Previous iterations on this branch focused on evaluating the multimodal capabilities of TowerVision, and the key updates are:

1. Improved post-processing functions for various benchmarks, including ALM-Bench, Commute, Kalidoscope, etc
2. Updated support for more languages in ALM-Bench.
3. Added support for the likelihood function across different multimodal models.
4. Introduced judges in Aya-Vision.

Additionally, we introduced a new benchmark called Blink.