Update TrainConfig optimizer hyperparameters

[jon-tow]

This PR provides the following updates based on observations from issue #53:

• Adds the previously unused weight_decay TrainConfig parameter to AccelerateRLModel optimizer initialization.
• Adds an opt_eps TrainConfig parameter to the AccelerateRLModel to allow users to specify an optimizer's epsilon value.
• Renames learning_rate_init and learning_rate_target to lr_init and lr_target TrainConfigs, respectively, for consistency with learning rate naming conventions in torch.optim algorithms.
• Removes unused lr_ramp_steps and lr_decay_steps TrainConfig parameters which look to be leftovers from the magicCARP lr scheduler.

[LouisCastricato]

@herbiebradley since this is for your issue.

[herbiebradley]

Looks good! Can we change the default epsilon to 1.0e-7 (default in Tensorflow) since the ablation paper found it was significantly better?

Also, does anyone know where the weight decay of 1.0e-6 comes from, since it is significantly lower than the default?

[maxreciprocate]

Also, does anyone know where the weight decay of 1.0e-6 comes from, since it is significantly lower than the default?

From here https://github.com/EleutherAI/magiCARP/blob/3537b4d7c98a4a384964856740a65731746e04fe/configs/base_config.yml#L18

[maxreciprocate]

lgtm!

[jon-tow]

@herbiebradley I intentionally left the default $\epsilon$ to 1.0e-8 as I wasn't 100% sure that the improvement would directly carry over to AdamW (we do not use vanilla Adam). I'd like to try some experiments before updating in another PR if that's okay.

[jon-tow]

cc #76 for awareness

[ayulockin]

Thanks for tagging the issue @jon-tow.