What happened: I tried to calculate a rolling mean or median, ignoring nan's

Minimal Complete Verifiable Example:

```python

da = xr.DataArray([1.0, 2, 3, np.nan, 5, 6]) da.rolling(dim_0 = 3).mean(skipna=True) <xarray.DataArray (dim_0: 6)> array([nan, nan, 2., nan, nan, nan]) Dimensions without coordinates: dim_0 ```

I expected this array to have have no nans. There is a simple workaround that does give the answer I was looking for:

```python

da.rolling(dim_0 = 3).construct("new").mean("new", skipna=True) <xarray.DataArray (dim_0: 6)> array([1. , 1.5, 2. , 2.5, 4. , 5.5]) Dimensions without coordinates: dim_0 ```

Anything else we need to know?: Love the project ;-)

Environment:

Hello all,

This is my first PR to a pydata project. This pull request is still very much a work in progress and I could really use your input on a couple of things.

• [x] Closes #2605
• [x] Tests added
• [x] Passes black . && mypy . && flake8

• [x] Fully documented, including whats-new.rst for all changes and api.rst for new API

• I moved the custom dask pad method into dask_array_compat to ensure backwards compatability to Dask versions that do not have dask.pad yet. We could chose to drop this support if we wanted to.

• I'm still in doubt about the function signature, numpy as dask use optional kwargs, but that kinda interferes with the **pad_width_kwargs. I chose the signature that I thought looked least awkward.
• The default behaviour of pad with mode=constant pads with NaN's converting the array to float in the process. This goes against the default behaviour of numpy as dask.
• How should the coordinates of a DataArray be padded? I chose default padding except for modes "edge", "reflect", "symmetric", "wrap".
• How should we handle inconsistencies between numpy.pad and Dask.pad, it turns out there are a couple 5303

Dataset.pad is coming up.

Hello xarray team

I was trying to implement functionality similar to Scikit Image's view_as_windows, but I was having a hard time with the boundary conditions. I understand this request is very similar to rolling with periodic boundary conditions #2007.

MCVE Code Sample

```python

arr = xr.DataArray(np.arange(4), dims=("x",)) arr.rolling(x=2).construct("roll_x", stride=2) <xarray.DataArray (x: 2, roll_x: 2)> array([[nan, 0.], [ 1., 2.]]) Dimensions without coordinates: x, roll_x ```

Expected Output

It would be nice to be able to easily get an output of: python <xarray.DataArray (x: 2, roll_x: 2)> array([[0., 1.], [2., 3.]]) Dimensions without coordinates: x, roll_x

Possible workarounds

With the upcoming features of pad (#3596) and rolling: periodic (#2011) it is actually not that hard to work around with:

```python arr.pad(x=(1,0)).rolling(x=2).construct("roll_x", stride=2).isel(x=slice(1,None))

or

arr.roll(x=-1).rolling(x=2, boundary="periodic").construct("roll_x", stride=2) ```

Idea

Would it be possible to add a keyword argument to either rolling() or construct() that expects an alignment of "left", "right" or "center". I understand that this could break interface copied from Pandas, but I hope it could be useful to others as well.