cache results of operations on data greta arrays

If you have two *data* greta arrays, and perform a (deterministic) operation on them, tensorflow (and therefore greta) will recalculate the output of each operation every time it is run*.

That's only rarely likely to be an issue. Hopefully users will do those computations first before converting to a greta array, either explicitly or implicitly. Though it's possible users will convert to greta arrays before doing long computations, in which case it would be kind of nice to handle that efficiently for them.

Operations that depend only on data could be identified from the DAG object (after defining the tensorflow graph), run once, and then their value stashed by greta, and used in the feed dict in the same way as the data will be.

---
* actually that's not quite true, since data is currently passed in as constants, but that will change soon because of #151