Use a proper datetime-like datatype in our output

[spencerahill]

Currently, the time dimension of aospy output is simply 'year' and an integer. E.g. data outputted from our tutorial:

In [40]: xr.open_dataset('~/Dropbox/py/aospy/aospy/examples/example-output/example_proj/example_model/example_run/precip_total/precip_total.ann.reg.ts.from_m
    ...: onthly_ts.example_model.example_run.0004-0006.nc')
Out[40]:
<xarray.Dataset>
Dimensions:              (year: 3)
Coordinates:
  * year                 (year) int32 1678 1679 1680
...

As we have discussed regarding custom reduction methods (#208), this indexing by year is too restrictive. I think we should replace it with a proper datetime object. This would enable using resample and groupby.

This is probably worth waiting until the cftime/netcdftime fixes upstream get incorporated. But I wanted to have this issue as a marker to make sure we don't forget this aspect of the output time types. @spencerkclark, any thoughts?

[spencerkclark]

I think this will largely be handled by adding functionality for custom reductions. Perhaps the default time reduction should be no reduction, as in #252 (which would maintain datetimes in the time coordinate in aospy output)?

Right now we always use groupby internally to convert data to a time series of annual means (this results in the 'year' coordinate of integer values), which is a totally valid reduction type, but perhaps not the most intuitive one for a default.