[NHM] Implement Kesten Processses with Smaller Memory Footprint on GPU

[HumphreyYang]

trafficstars

As discussed in #37.

From @jstac: I feel there might be a useful middle ground here. For example, use lax.scan to run forward 100 time periods and then call that 5 or 10 times in an outer loop in pure Python.

[jstac]

Thanks @HumphreyYang . Applicable to Kesten process lecture, inventory lecture, wealth dynamics lecture.