consider infinite horizon problems

[marcdelabarrera]

Right now the package only solves finite horizon problems. Each training loop has a well-defined length T and the policy function depends on t. Consider a model with no discounting where in each period, with probability $1-\beta$ the agent dies and utility is 0 from then on. Then in expectation, that should be as if the agent is discounting time at rate $\beta$. Simulate until death and do not allow the policy function to depend on t.