Add CHELSA-ERA5 tiled climatologies.

[juseg]

• [x] Add CERA5TiledAggregator class.
• [x] Add 'cera5' atmospheric data source.
• [ ] Add 'cera5' elevation data source.
• [ ] Consider renaming coords as in 'cw5e5'.
• [ ] Update docstrings and whatsnew.rst.

Close #96.

[juseg]

This works. The data re-projection is blazing-fast in comparison to working with the geotiffs (as evidenced using a dask.diagnostic.ProgressBar(). However, the overall atmosphere() call time is barely faster. I think this is due to the large overhead of reading metadata from 864 netCDF tile files.

Now I assume CHELSA-W5E5 suffers the same problem. I need to think about that. Ideas for the way forward:

• Don't tile, but use global, spatially-chunked netCDF.
• Use larger tiles, e.g. 90x90 will result in 8 tiles, 60 files.
• Keep tile size, but aggregate months in 72 tiles, 72 files.

Probably a (yet another) separate issue. I like the third solution because (hypothetical, some-day) end-users would normally use data for all months on a given region. It would also pave the road towards smarter aggregators opening (or even downloading) only necessary files.

[juseg]

Benchmarks on an preparing an Alps domain dataset. Averages of 5 calls, using dask.diagnostics for reprojections time.

• Old 'chelsa' 12 global geotiffs: reproj 2 x 5 secs, total 17 secs.
• New 'cw5e5' 864 monthly tiles: reproj 2 x 120 ms, total 9 secs.
• New 'cera5' 72 yearlong tiles: reproj 2 x 300ms, total 6 secs.

Total times still include re-projecting the global netCDF topography, which seems to take most of 6 secs in the last case.