docs(gradmax): rewrite GradMax algorithm page

[maxencelebaron]

This PR adds the documentation page for the GradMax algorithm.

Content:

• Introduction: GradMax focuses solely on the "how to grow" question, initializing new neurons by maximizing gradient norms rather than greedily minimizing the loss
• Theory: general optimization problem
• Experiments: two result tables on CIFAR-10/100 and ImageNet comparing GradMax against random initialization, Firefly, and baselines

Files changed:

• docs/algorithms/gradmax.rst: new algorithm page
• docs/_static/gradmax.png: figure illustrating neuron addition in GradMax

[TheoRudkiewicz]

Thank you for this PR. Here are a few comments:

• you should consider that people already red the introduction and in particular https://growingnet.github.io/growing_wiki/overview/neuron_addition_problem.html. As a consequence it would be better if you use coherent notations (Psi / Omega)
• Even if it's not the goal of GradMax it is important to report how they solve When, Where and How many to be able to compare with other methods.
• If you could give at least one-line explanation about "Random" (I think the paper is unclear, hopefully the code is clearer). In particular, if it is a Kaiming style init, what is the fan-in size considered ?

[TheoRudkiewicz]

I think this is false (it's a misstake of the paper). It should add the hypotestis of orthogonality of the different component.

[maxencelebaron]

Thanks for the review! If I understand correctly, without an orthogonality constraint on the columns of $W_{\ell+1}^{\text{new}}$, the SVD solution (top-k left singular vectors) is not the actual optimum. Is that what you mean?

[TheoRudkiewicz]

I a am bit skeptical about the weight decay value.

[maxencelebaron]

Yes, you're right. That's a mistake. The correct value is 2e-4.