Support segmented scans

Is your feature request related to a problem?

It is pretty common to want to run cumsum and have the sum reset when a boolean flag array is 1. This is so common it has its own Wikipedia page and is discussed in Blelloch (1993) (Section 1.5)

Here's a real example of someone trying to implement it in a fairly roundabout way.

time_cumsum = cube.cumsum(dim = 'time')
cumsum = time_cumsum - time_cumsum.where(cube== 0).ffill(dim = 'time').fillna(0)

We have a few options to implement it:

1. We could introduce a new method DataArray.segmented_scan(flags, op="sum") or a new class DataArray.segment.cumsum()? A dask/cubed friendly version that does all of this in a single scan should be fairly straightforward to write (and similar to our ffill, bfill wrappers).

2. In a way this generalizes resample and it just struck me that the example above could be written as the following, which should be OK once flox adds scans

   group_idx = (cube == 0).cumsum('time')
   cubed.groupby(group_idx).cumsum()
   

   1. We could use our new Grouper functionality to expose a "flag" grouper that hides the group_idx = (cube == 0).cumsum('time') line.

My concern with (2) and (2.i) is that they are not at all obvious for most of our userbase.

The issue does not name implementation files or tests. Start by comparing the proposed DataArray.segmented_scan API with the flag Grouper approach, and review the existing cumsum and ffill wrappers mentioned in the issue. Done requires a decided API and a dask/cubed-friendly segmented scan.