netcdf save rehash : beginnings of append support

[pp-mo]

WIP : DO NOT MERGE
Post-refactor (#51) code, beginning work on implementing a netcdf-append strategy

Status: highly preliminary (!) :

• the _offset_keys routine may not now be necessary
  see http://docs.dask.org/en/latest/array-api.html#dask.array.store
  -- use of the regions key may be able to replace this.
• the actual "append" operation is really very simple ...
• BUT the 'validity checking' code in the _prepare_append method is huge + still untested
  it has become a monster !!
• the 'netcdf data representation' classes introduced in netcdf save rehash : working refactor (to support future append) #51 now look doubtful to me
  • use of '_SlotsHolder' may be excessive : simpler to define classes directly ?
  • or replace with xarray objects (???)
• the use of attached "tracking" properties in the 'label_links' routine (inner routine to _prepare_append) seems inappropriate
  • almost certainly clearer to store relational info in separate mapping objects (and maybe even simpler)

[pp-mo]

Regarding "replace with xarray objects (?)",
Maybe this whole effort would sit better in the xarray project (where they already support appending writes, though at present just for adding new variables, not extending existing ones AFAICT)
- in that case, we should approach it in Iris by improving xarray integration instead : provide save-to-xarray as alternative to save-to-file.

Presumably that should mean loading, too.

Interoperability is currently provided in xarray (to_iris + from_iris),
but of course that doesn't implement all the CF logic that Iris can.
If Iris has a more sophisticated view of CF data, then it makes sense to define xarray interoperability in Iris, instead ??