feat: Dormand-Prince RK45 adaptive ODE solver

[bwyard]

Context

rk4_step uses fixed step size. Stiff systems and rapidly-changing dynamics need adaptive step-size control to maintain accuracy without wasting evaluations.

Proposed API

pub fn rk45_adaptive(
    state: f32, t0: f32, t_end: f32, dt_initial: f32, tol: f32,
    f: impl Fn(f32, f32) -> f32,
) -> (f32, f32, u32)  // (final_state, final_time, steps_taken)

Dormand-Prince method with embedded 4th/5th order error estimate. Step size adjusted by dt *= 0.9 * (tol/error)^0.2.

Thesis alignment

• Pure: (state, t0, t_end, dt, tol, f) -> (state, t, steps)
• ADVANCE-EXCEPTION: while loop bounded by max_steps
• No &mut in public API

Tests needed

• dy/dt = -y gives exp(-t) within tolerance
• Oscillator energy conservation
• Step count is reasonable (fewer than fixed-step for same accuracy)
• Deterministic