Learning the Linear Quadratic Regulator from Nonlinear Observations