Likelihood Training of Schr\"odinger Bridge using Forward-Backward SDEs Theory