ADOPT, MARS, Prodigy

[Andron00e]

this pr brings three more methods.
i think there might be an issue with MARS (its mars-shampoo version), ill check today.

comment: for MARS, i changed train_step behaviour a bit: ref

since were using only approximate scheme of MARS, we should only track the past gradient in order to compute g_curr - g_prev, therefore i use this in the train_step:

if optimizer.__class__.__name__ == "MARS":
     optimizer.zero_grad(set_to_none=True) 
     optimizer.update_last_grad()
else:
    optimizer.zero_grad()