Deprecate / remove support for n5 attribute mapping "gimmick"

[bogovicj]

We should seriously consider deprecating or removing support for the n5-attribute-mapping "gimmick" .

For example that

n5.getAttribute(dataset, "dimension", long[].class);

returns the zarr's shape attribute, when n5 points to a zarr dataset.

Reasons:

1. There exist other was to get the same information that are more reliable and more easily maintained. (see additional details below)
2. It adds substantial maintenance burden without a clear (to me) purpose.
3. It is unclear how it should behave in the future as zarr changes ( for example, zarr 3 discussed briefly here)

Reason 1

There exist other ways to get the information provided by this feature.

For example, the "dimensions" example above is equivalent to:

n5.getDatasetAttributes(dataset).getDimensions();

and should be preferred to the method above.
It will be more reliable for example, if and when there is a clash with the mapped attribute name (say, if `"dimensions" appears in .zattrs).

Further, removing "official" support does not prevent implementations from doing this themselves if they choose to.
Implementations can check if an N5Reader is a ZarrReader and getAttribute for "shape". Alternatively, an implementation
could use metadata translation to virtually add those attributes if desired.

@axtimwalde
@tpietzsch
@cmhulbert