Bug: Spin and charge get lost in optimizers

[niklashoelter]

Related to #253 #277 , initially adressed in #373

The passing of the spin and charge parameters from ASE to the Torchsim SimState works.
When passing this SimState through a single model iteration (tested with fairchem), also everything works fine.

The problem occurs in optimization runs - where the spin/charge information as far as I can see gets lost in the init_fn of the optimizer (I observed this using fire, but probably gradient_descent faces the same problem).
Suggested fix: Copy spin and charge from the SimState argument to the common_args dict in optimizers/fire.py init_fire().

Definitely need to check if the problem also persists in the step_fn, as attributes are also passed here - I am unsure about this.

[thomasloux]

Good catch, in most of ìnit_fn`, it passes under the hood the arguments to a class that inherits from SimState. It's possible that we need to modify most of such init functions.
You need to modify this dict to add spin and charge:
https://github.com/TorchSim/torch-sim/blob/main/torch_sim/optimizers/gradient_descent.py#L54