Cannot pickle cubes with GRIB data

[pp-mo]

(#201 moves the grib-specific part of the iris 'test-pickle' from iris into iris-grib.
In principle, a cube should be picklable, but at present this doesn't work with grib data.
In fact a GribMessage itself cannot be successfully pickled + brought back.
So that needs fixing.
)

Actually, I think this issue was fixed long ago and obsolete
But now I think it is back ! See comment

Update again

I was wrong, this problem never went away.
Regardless, I do think this is still an open problem.

As posted elsewhere ...

Isn't the whole point of our interest in pickling that it is essential to the use of dask.distributed ?

In which case...

• the fact that you can't successfully reconstruct a pickled cube derived from a grib load is still very much a problem.
• No idea how important though ... or what it takes to fix ... but I believe this has proved awkward in the past
• possibly something to do with the way ecCodes (and gribapi before it) use "handles" for messages ?
  • -- so presumably needs (at least) a getstate/setstate solution -- or a fix, since DataProxy at least already has one
• possibly just needs a smarter definition of 'equality' to fix it

HINT:

I think latest dask problems seen in Iris have highlighted problems specifically with "__slots__",
so that might affect DataProxy ??
Note that _DataProxy doesn't seem fit the bill for that error message -- but possibly _DataProxy.recreate_raw, which is a _MessageLocation does ?