Add dask.blockwise.fuse_roots optimization

[mrocklin]

This fuses data generation or access graphs into blockwise after
converting everything to low level task graphs

cc @TomAugspurger we were just talking about this

[mrocklin]

Tests pass. I'm going to wait until next week to make sure that the 2.7.0 release is smooth, and then I'll merge this in if there are no objections.

[mrocklin]

@TomAugspurger @rabernat this is in. It should help avoid communication in situations where multiple arrays are loaded and then combined together as in z = x + y. I believe that this was causing you some heartache with excessive memory use. Some of your workloads may now be smoother with this in.

[dcherian]

Exciting! Thanks @mrocklin