Accuracy in simulations

[mfalt]

Since changing to continuous simulations we lose accuracy in lsim compared to the previous zoh method when the input is already zoh. (see #118 (comment))
Step was, for example, exact almost up to machine precision before. We should do at least one of the following

• Default step to zoh when the t input is equidistant (this was the previous behaviour)
• Allow time responses to accept a tolerance argument that is sent to the numerical solver
• Lower the default tolerance in call to the numerical solver