Quick attempt at

1. CF-compliant encoding of Variables wrapping pd.IntervalArray totwo variables.
2. decoding of CF-compliant bounds variables (e.g.time_bounds) to an IntervalArray. For this to be useful when a central valuetime is present requires that we droptime, create the intervalindex fortime_bounds and then assign that astime.

Alternatively we might not want to record the "central value" on encode. Then we could simply have a 1D Variable(pd.IntervalArray) -> 2D Variable encoding pipeline with attributes:

1. closed: 'left'. which side of the interval is closed?
2. xarray_dtype: 'pandas_interval'. tells Xarray to decode to thepd.Interval dtype
3. bounds_dim_name: 'bounds' that identifies the dimension with size-2 (left & right edges)

This is not CF-compliant but would let usclose#2847