Speed up Dataset._construct_dataarray

[dcherian]

• Tests added
• Passes isort . && black . && mypy . && flake8
• User visible changes (including notable bug fixes) are documented in whats-new.rst

Significantly speeds up _construct_dataarray by iterating over ._coord_names instead of .coords. This avoids unnecessarily constructing a DatasetCoordinates object and massively speeds up repr construction for datasets with large numbers of variables.

Construct a 2000 variable dataset

importnumpyasnpimportxarrayasxra=np.arange(0, 2000)
b=np.core.defchararray.add("long_variable_name", a.astype(str))
coords=dict(time=np.array([0, 1]))
data_vars=dict()
forvinb:
data_vars[v] =xr.DataArray(
name=v,
data=np.array([3, 4]),
dims=["time"],
coords=coords
)
ds0=xr.Dataset(data_vars)

Before:

%timeit ds0['long_variable_name1999']
%timeit ds0.__repr__()
1.33 ms ± 23 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
2.66 s ± 52.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

After:

%timeit ds0['long_variable_name1999']
%timeit ds0.__repr__()
10.5 µs ± 203 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
84.2 ms ± 1.28 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

[dcherian]

Merging. This logic could be updated if we convert _coord_names to a list or OrderedSet.