Bug with comparing ManifestArrays using __eq__ which have fill_value=NaN

[TomNicholas]

I'm stuck on why this assert_identical on the datasets doesn't pass. It complains that all the coordinates are different, but every attribute I can think to check looks identical 😖

Originally posted by @TomNicholas in #349 (comment)

I found out what the problem is - two ManifestArrays with identical .metadata will compare not equal if they both have NaN for a fill_value. This is because the __eq__ check introspects deeper until it finds e.g. the np.float32(nan) type, but

In [4]: bool(np.float32('nan') == np.float32('nan'))
Out[4]: False

(See https://stackoverflow.com/a/10059796 for why numpy NaNs behave like this.)

The solution needs to be to actually check two Metadata classes are __eq__ with dedicated code, not just trusting the python dataclasses' automatically-generated __eq__ method to fi it.

I would ideally like to move that fix upstream into zarr-python.