Sampling posterior predictive from shifted Lognormal

[junpenglao]

You need draw_values to fetch concrete value instead of tensor from the graph: https://github.com/pymc-devs/pymc3/blob/master/pymc3/distributions/distribution.py#L524