Gravar-mail: A Bayesian proportional hazards regression model with non-ignorably missing time-varying covariates