Tolerance has to do with the numerical algorithm used to find the solution to the objective function underlying lasso regression, in this case coordinate descent. It could most affect accuracy but not in a way you can predict. You would want to adjust it if your model is not converging within the iteration limit that is set. The objective function describes a space with local maxima and minima, and altering tolerance could help you more tightly find one of those or move past one and on to another but that does not guarantee better accuracy. "No free lunch"
eps I am less clear on. It states in the doc that this is the ratio of alpha min to alpha max, and alpha is the coefficient in front of the penalization term. So epsilon seems like it would limit the path of alpha values that are being tested by the coordinate descent (optimization) algorithm. I think increasing epsilon could improve your model but at a sharp cost in computation time.