Replace `OptaxOptimistixProximalGradient` with a port of `jaxopt.ProximalGradient`

[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 is not as theoretically sound as jaxopt.ProximalGradient which implements the FISTA algorithm.

To be able to eventually stop vendoring (and drop support for) JAXopt, we decided to add a port of JAXopt's ProximalGradient that adheres to Optimistix's interface, and replace OptaxOptimistixProximalGradient with this when using the Optimistix backend.

[bagibence]

Similarly, to support the original Nesterov acceleration, adapt this to create a port of GradientDescent.

[bagibence]

When this is done:

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