Reduce memory usage: zero-copy

[aufdenkampe]

In @sjordan29's exploration of performance issues in coupling TSM to a test case of ClearWater-riverine, we suspected that we might be running out of memory. This corresponded to our performance profiling findings described in #57 (comment) where a substantial amount of time was spent in xarray merge/copy/init functions.

Zero-copy

We want to move as much as possible to a zero-copy approach when running (and coupling) these models. Remember that xarray.DataArray is really just a numpy.ndarray under the hood, so we should use pointers to that single source of data (rather than copying) as much as possible.

Zero-copy approaches complement the pre-initialized array approach described in:

• Improve timestep performance further: numba or not? #57 (comment)

[aufdenkampe]

I just edited this issue title and description to split out the following into separate issues:

• Reduce memory usage: lazy writing to dask.array #69
• Add memory profiling #70