Add finite difference and sample options to LBFGS

[nikitakuklev]

Added with_grad, which is passed down to scipy method. BoTorch automatically uses with torch.no_grad(), so we don't need to disable gradients ourselves. Using finite differences, non-differentiable operations can be handled reasonably.

Re-added n_raw_samples - this is supposed to be a helpful feature to ensure that restarts begin on good points. It was removed awhile back, not sure why.