Xarray: Feature Request: Efficient rolling with strides

[niowniow]

Xarray is facing the same issues in its current rolling implementation (DataArrayRolling and DatasetRolling) as described in this pandas issue. Namely, the construct methods stride parameter is applied after the rolling is computed. Technically, we are computing more than we would need to because we partially throwing it away due to striding.

In PR #3607 the issue is solved for the ...Rolling's __iter__ function but not for the construct, reduce and _bottleneck_reduce methods.
Since the way Xarray's rolling is implemented relies on numpy, we could introduce a sliding window function as described here.

Any opinions?

[niowniow]

Previous enhancement requests asking for a stride argument to rolling: https://github.com/pandas-dev/pandas/issues/15354, https://github.com/pandas-dev/pandas/issues/22976, https://github.com/pandas-dev/pandas/issues/27654#issue-474416717, https://github.com/dask/dask/issues/4659, https://github.com/numpy/numpy/issues/7753

_Originally posted by @pilkibun in https://github.com/pandas-dev/pandas/issues/26959#issuecomment-511233955_