Allow ReadInternalStore to be pickled

[braingram]

To use a zarr array read from an asdf file (with chunks stored as internal ASDF blocks) with dask distributed processing the store that provides the chunk data must be pickleable.

This PR makes ReadInternalStore pickleable (and re-pickleable as dask may transfer chunks between workers).

The following example was used for prototyping:

import os

import asdf
import dask.distributed
import dask.array
import zarr
import numpy as np

import asdf_zarr.storage


if __name__ == '__main__':
    fn = 'big_test.asdf'

    if not os.path.exists(fn):
        shape = (17000, 37000)  # 600 MB
        a = np.arange(np.product(shape), dtype='uint8').reshape(shape)
        z = zarr.array(a, chunks=(1000, 1000), compressor=None)
        af = asdf.AsdfFile()
        af['z'] = asdf_zarr.storage.to_internal(z)
        af.write_to(fn)


    client = dask.distributed.Client()
    print(f"client at {client.dashboard_link}")


    with asdf.open(fn) as af:
        z = af['z']
        d = dask.array.from_zarr(z)
        input("press enter to compute sum")
        print("Computing sum...")
        s = d.sum().compute()
        print(f"\tsum={s}")
        input("press enter to exit")

The above example generates a large (17k x 37k) image, saves it to an ASDF file as a chunked array, loads the saved file, converts the chunked array to a dask array and computes the sum (using dask distributed). The screenshot below is of a single run showing the computation distributed across 10 workers.

As it's likely that users will want to use dask we should consider options for testing dask compatibility (and specifically dask distributed compatibility) in the CI.

[Cadair]

🥳

[braingram]

The private _attr argument is used here (as is needed for ndarray as noted in the code). Perhaps this can inform what a "low level" block API would look like. By using the offset and the uri this will not work (and likely produce junk data) after an update of the backing ASDF file. Minimally we should document this danger and it may be one better to check the file mode and not allow pickling when the file is opened 'rw'.

[braingram]

This is another instance of using non-public asdf api. I would be happy to see 3.x include at least a provisional introduce of a lower level api that would include a function that could replace this (that loads a block from an ASDF file without parsing the tree, etc).

[braingram]

This could be improved by changing InternalStore to not rely on code in __init__. I will open a follow-up issue.