Adds automatic initial timestep selection following Hairer-Norsett-Wanner.

• Simulation.estimate_initial_step() forms a first guess from the scaled norms of the state and its derivative, then refines it with a second-derivative estimate from an explicit Euler probe (a pure evaluation, no actual timestep / no events). Scaling uses the engine error tolerances.
• Blocks expose their state derivative through a new Block.derivative(t) accessor: the base implementation covers blocks with a dynamic operator (ODE, DynamicalSystem, LTI, and the DAE blocks via their reduced operator); Integrator overrides it; stateless blocks return None and are skipped.
• Passing dt=None to the simulation triggers the estimate at the start of the run; otherwise behaviour is unchanged.

Tests cover the derivative accessor, step scaling with the dynamics, state restoration, and an end-to-end auto-dt run.