Reconcile Sampler protocol with `model_wrapper.step()` arguments required by Protpardelle

[marcuscollins]

The following change in edm.py in https://github.com/diff-use/sampleworks/pull/274/changes#diff-db2bf8d2d9093b15fb1f471360efc972f4b0ec228b42c6e91d27e5f7081320f7R428

        # TODO testing adding eps to signature for use with Protpardelle-1c, if successful,
        #   I need to modify the Protocol itself. @Michael Anzuoni
        with torch.set_grad_enabled(allow_gradients):
            x_hat_0 = model_wrapper.step(noisy_state, t_hat, eps, features=features)

technically breaks our Sampler protocol by passing an extra argument (here, eps, which is a noise level). It was introduced into the branch mdc/add-propardelle to properly handle frame transformations (since we should transform before adding noise in Protpardelle-1c.)

I (marcuscollins) don't think this will break any of the actual code, since the model wrappers all include a * for extra positional arguments to their step() methods. The question is how to handle it in the Protocol, when some arguments are required and others are not.

If PR https://github.com/diff-use/sampleworks/pull/274/changes isn't yet merged, feel free to add any changes there.

[k-chrispens]

See PR comments. I believe this should be closed, as it was already addressed by #267

[k-chrispens]

Also, just to be clear - eps is not a noise level. eps is the noise tensor to be added at a given step.