Revisit methodology

[nocollier]

The (Collier2018](https://doi.org/10.1029/2018MS001354) methodology normalizes errors by the standard deviation of the reference variable:

s(x) = exp( - bias(x) / std( ref(t, x) ) )

This worked reasonably well for land, but leads to not-helpful error maps for ocean where the variability is comparatively small. We should rethink this in the context of ocean benchmarking. One alternative would be to use regional quantiles over ocean basins as we have proposed and implemented in ILAMB.