disallow numpy scalars for metadata.fill_value

[d-v-b]

Right now the fill_value attribute of our metadata classes is a numpy scalar, or None, or a string (i think). This is problematic for comparing two metadata instances -- if the fill value is np.nan, then np.nan == np.nan is always false, even when the two values are in fact the same.

We could address this by disallowing numpy scalars for the fill_value attribute. Instead, we could use a custom scalar wrapper class that handles NaN correctly, or use a JSON-serializable form of the scalar, e.g. a string in the case of np.nan.

see also: #2929

[d-v-b]

an alternative approach: we allow scalars, but we override __eq__ to compare the JSON form of the metadata.

This seems better to me than requiring that the fill_value attribute be JSON-serialized. If we pre-JSONify the fill value, a string array with a fill value that is the string "nan" would falsely appear to have the same fill value as a float array with a np.NaN fill value.