Still pretty complicated. Probably need some helpers in order to test the more interesting stuff. Also would like to hypothesis?

opening asset table probably works; reads still todo. a few missteps along the way:
* `np.vectorize` on `asset_table_to_reader_and_window` is segfaulting. May have to do with all the extra arguments? Don't care to gdb it; just moved the vectorization into the function
* Blockwise wants to fuse the reader table to the reads. I'm not sure it's good that blockwise optimization tries to fuse such a large expansion, but probably kinda makes sense. Unfortunately the way to disable it (annotations) feels like a little bit of a hack

Passing dtype and fill_value into `fetch_raster_window`, instead of storing them in the reader table, simplifies things a lot. Also lets us get rid of those fields on the reader protocol.

small steps!

[PEP 585](https://www.python.org/dev/peps/pep-0585/) is only available in 3.9. We're still supporting 3.8 for now; planetary computer uses it, as do probably many others