Slightly optimize computation for RigidContacts model

[flferretti]

This PR slightly enhances the performances of the RigidContacts model by using the pseudoinverse instead of the least-squares method to compute the impact velocity, and removes some unnecessary context managers and assertions

---

📚 Documentation preview 📚: https://jaxsim--325.org.readthedocs.build//325/

[CarlottaSartore]

We checked with @flferretti and the time needed to perform jnp.linalg.pinv and jnp.linalg.lstsq are comparable but since lstq is more stable for ill conditioned matrix , it could have sense to keep the original implementation.

Maybe we can close this PR then @flferretti ?

[xela-95]

Thanks @flferretti ! I left a comment

[xela-95]

We checked with @flferretti and the time needed to perform jnp.linalg.pinv and jnp.linalg.lstsq are comparable but since lstq is more stable for ill conditioned matrix , it could have sense to keep the original implementation.

Maybe we can close this PR then @flferretti ?

Sorry @CarlottaSartore I skipped your comment! Mmmh yes, then maybe we can close this for now. Do you agree @flferretti ?

[flferretti]

Sorry @CarlottaSartore I skipped your comment! Mmmh yes, then maybe we can close this for now. Do you agree @flferretti ?

Maybe we can close this PR then @flferretti ?

Sure! I also removed a switch of the velocity representation and refactored a bit the function. If you agree I can restore the jnp.linalg.solve and keep the rest as it is