Add port of FISTA and NAG from JAXopt

[bagibence]

As Optimistix doesn't implement ProximalGradient, this is currently set to use a custom implementation in OptaxOptimistixProximalGradient based on Optax's SGD with linesearch followed by the proximal operator.
This seems to work in practice, but doesn't support the original Nesterov acceleration and is not as theoretically sound as jaxopt.ProximalGradient which implements the FISTA algorithm.

This PR adds OptimistixFISTA, a port of JAXopt's ProximalGradient that adheres to Optimistix's interface, and replaces OptaxOptimistixProximalGradient with this when using the Optimistix backend.

Similarly, to support the original Nesterov acceleration in gradient descent, a short modification of this as OptimistixNAG is added. NAG stands for Nesterov's Accelerated Gradient.

Additional changes:

• Remove OptimistixOptaxProximalGradient
• Remove stateful_scale_by_learning_rate (and related code) which was only needed for OptimistixOptaxProximalGradient
• Switch OptimistixOptaxGradientDescent to use optax.scale_by_backtracking_linesearch

OptimistixOptaxGradientDescent is kept but is replaced in the tests. Ideally, it would be tested the same way as OptimistixNAG.
@BalzaniEdoardo, @billbrod is that fine with you, or is a way to include this in the tests without unnecessarily rerunning others required?

Fixes #380

[BalzaniEdoardo]

I would rename this as _fista_linesearch.py or something like that. True that the algorithm is inspired by the jaxopt implementation but you did change enough of it to match the optimistix API that we can omit the fact that it is a port.

[bagibence]

I renamed it to _fista.py and rephrased the module's docstring, keeping "Adapted from JAXopt." to credit them.

[BalzaniEdoardo]

If I understand correctly, the default is using OptimistixNAG but we could still use the Optax GD if one wants to.

One way to test this is to use unittest mock to mock the solver registry in the test:

from unittest.mock import patch


# In your test
with patch.dict('nemos.solvers.solver_registry', 
                {'GradientDescent': nmo.solvers.OptimistixOptaxGradientDescent}):
    model = Model(solver_name='GradientDescent2', ...)
    model.instantiate_solver()
    
    assert isinstance(model._solver, nmo.solvers.OptimistixOptaxGradientDescent)

[bagibence]

The problem with that is that it would have to be done for every test that uses gradient descent.
I added a solution where only tests referencing "GradientDescent" explicitly are rerun with the OptimistixOptaxGradientDescent implementation and added it as a step to the CI job for the Optimistix backend.

[bagibence]

Changed this so that it is now controlled by pytest -o override_solver=GradientDescent:OptimistixOptaxGradientDescent and testing both implementations is done in the same tox environment.