Feature: Add Relaxed Archetypal (RA) implementations for TopK and Jump SAEs

[SanBast]

Description: This PR implements RATopKSAE and RAJumpSAE. These classes extend the existing TopK and Jump SAE architectures by integrating the existing RelaxedArchetypalDictionary class in the provided original repo.

Changes:

1. Created overcomplete/sae/ra_sae.py containing: RATopKSAE: Inherits from TopKSAE and RAJumpSAE: Inherits from JumpSAE.
2. Updated overcomplete/sae/init.py to expose these new classes.

Motivation: I'm a huge fan of this repo, and I'm using RA SAE for my studies in medical XAI and segmentation. I wanted to create this feature in order to be used off-the-shelf :)

Testing: I have tested instantiation and a forward pass for both classes to ensure shape consistency with the parent classes

[fel-thomas]

Hi ;)
Thank a lot yes indeed good idea.

I'll add some tests, remove the #todo and rebase that today or tomorrow, and thanks for the typo catch "bandwidth" you are correct!

[fel-thomas]

Added 60+ tests, all ok, LGTM and thank again !
I will write a small notebook in the next days :)