Separate the hybrid weights by variables/domains/locations

[guillaumevernieres]

Description

Grouping the localization operators by domains (ocean and ice) is not working. The initialization of bump/nicas seems to be fine, but the envar produces nans. I'm pausing that development for now, the idea was to remove the cross-correlation between the ocean and seaice. This can be done differently.

Solution

1. Back to duplicated localization with only one group. This will generate covariances across ocean and seaice.
2. This isn't exactly what I had in mind, but since the univariate assimilation of ice concentration has been producing reasonable results so far, I'm going to "turn-off" or reduce the impact of the ice ensemble by providing spatially varying and variable dependent weights. See here: https://github.com/JCSDA-internal/oops/blob/1d41d55f510957dfb4102b4c3be69189927d9b61/src/oops/base/HybridCovariance.h#L89