Expose xarray's h5py serialization capabilites as public API?

Xarray has a magic ability to serialize h5py datasets. We should expose this somehow and allow it to be used outside of xarray.

Consider the following example:

import s3fs
import h5py
import dask.array as dsa
import xarray as xr
import cloudpickle

url = 'noaa-goes16/ABI-L2-RRQPEF/2020/001/00/OR_ABI-L2-RRQPEF-M6_G16_s20200010000216_e20200010009524_c20200010010034.nc'
fs = s3fs.S3FileSystem(anon=True)
f = fs.open(url)
ds = h5py.File(f, mode='r')
data = dsa.from_array(ds['RRQPE'])
_ = cloudpickle.dumps(data)

This raises TypeError: h5py objects cannot be pickled.

However, if I read the file with xarray...

ds = xr.open_dataset(f, chunks={})
data = ds['RRQPE'].data
_ = cloudpickle.dumps(data)

It works just fine. This has come up in several places (e.g. https://github.com/dask/s3fs/issues/337, https://github.com/dask/distributed/issues/2787).

It seems like the ability to pickle these arrays is broadly useful, beyond xarray.

1. How does our magic work?
2. What would it look like to break this magic out and expose it as public API (or inside another package)

Start by reproducing the h5py pickling failure in the issue examples and tracing how xarray makes the equivalent dask-backed array serializable. Determine where that behavior currently lives, then assess a public API or separate-package design and define the required tests and documentation before implementation.