Special treatment of NaN in JSON user attributes

[meggart]

This is a problem that showed up when testing the Zarr Julia interface. Currently nan-values in user-attribute values are not converted to strings (see example below), which does not adhere to JSON specs and causes the Julia JSON parser to fail. So I would suggest to do something analogous to the fill_value entry in .zarray, i.e. to either forbid nan in zarr user attributes or convert them to strings to make the resulting JSON valid.

import zarr
import numpy
z1 = zarr.open('attrtest')
z1.attrs['att1']=numpy.nan

with open('attrtest/.zattrs') as myfile:
    print(myfile.read())

[manzt]

Just want to keep track of this: zarr-developers/community#37 (comment)

NaN as fill_value seem to be converted to strings, but more care needs to be taken when parsing .zattrs in other languages since python will not but NaN into strings by default, and NaN is not valid JSON.

[Carreau]

This is also related to zarr-developers/zarr-specs#81, the stdlib json does not respect the json spec, and we should decide what to do about this.

[jhamman]

This was included in the v3 spec: https://zarr-specs.readthedocs.io/en/latest/v3/core/v3.0.html#fill-value