Learning Exponential Families from Truncated Samples