Implement `skipna` in xr.quantile for speedup

[aaronspring]

xr.quantile uses np.nanquantile which is slower than np.quantile but only needed when ignoring nans is needed. Adding skipna as kwarg would lead to a speedup for many use-cases.

MCVE Code Sample

np.quantile is much faster than np.nanquantile

control = xr.DataArray(np.random.random((50,256,192)),dims=['time','x','y'])
%time _ = control.quantile(dim='time',q=q)
CPU times: user 4.14 s, sys: 61.4 ms, total: 4.2 s
Wall time: 4.3 s


%time _ = np.quantile(control,q,axis=0)
CPU times: user 47.1 ms, sys: 4.27 ms, total: 51.4 ms
Wall time: 52.6 ms

%time _ = np.nanquantile(control,q,axis=0)
CPU times: user 3.18 s, sys: 21.4 ms, total: 3.2 s
Wall time: 3.22 s

Expected Output

faster xr.quantile:

%time _ = control.quantile(dim='time',q=q)
CPU times: user 4.95 s, sys: 34.3 ms, total: 4.98 s
Wall time: 5.88 s

%time _ = control.quantile(dim='time',q=q, skipna=False)
CPU times: user 85.3 ms, sys: 16.7 ms, total: 102 ms
Wall time: 127 ms

Problem Description

np.nanquantile not always needed

Versions

Output of `xr.show_versions()`

xr=0.15.1

[max-sixty]

Great -- we'd take a PR for this. Cheers @aaronspring