Provably efficient RL with Rich Observations via Latent State Decoding