Plan for high level graph optimizations

Now that we have high level graphs in our collections, we can do some more complex optimizations. We've done a little bit of this with blockwise fusion, read_parquet and column selection, and root fusion, all of which have had really positive effects.

There is likely more that we can do here. Some thoughts ...

- Fuse column selection with other kinds of data access, like read_csv or ORC
- Pass slicing through some blockwise operations in Dask array (a long-held request by @shoyer)
- Optimize the subgraph callables in blockwise with Numba (or something else) for dask array to avoid memory copies and maybe reduce serialization time
- Swap around Joins and filters and column access
- ...

It would be nice to have a current maintainer consider these options, guess how long they would take, and see what makes sense to do short term