global flag in data stage shoots up memory usage

[raehik]

Setting --global_ 1 when running cmip26.py balloons memory usage. @arthurBarthe offers that an xarray op used there might have changed behaviour.

The readme uses --global_ 1 in its example MLflow-wrapped command. That should probably be changed or come with a warning. It's set to 0 by default in cmip26.py.

[arthurBarthe]

Probably a minor issue at the moment. The global parameter is set to 1 so that when we use the full data, it is "periodized" along the longitudinal dimension prior to the data processing so that the convolution operations do not incur data loss along that dimension. Better in order to have nicer plots (without a missing vertical band).