Set resources for dataloading, training and prediction

[SarahAlidoost]

closes #44

This PR:

• adds a util function to set the resources and shows how to set them in the example notebook
• fixes a few minor things
• adds a comparison between baseline (monthly average) and model results on residuals via plots

[CLAassistant]

All committers have signed the CLA.

[SarahAlidoost]

@meiertgrootes and @rogerkuou when working on this PR, I found that the land mask from era5 does not mask the large water bodies, for example see the plots below, the white pixels are related to Lake Victoria in Africa. These large bodies are not available in "averaged" daily sst data. Something to be discussed in our meeting (data checkup).

Blue regions are water (sea, lakes, rivers,...), green are land.

White pixels (in land region) are nan in "averaged" daily sst data.

[SarahAlidoost]

@meiertgrootes the notebook uses residuals as target and compares the model output with the baseline (averaged daily data).

[meiertgrootes]

The PR looks good! All changes should be very useful. I've made acomment suggesting possible additional diagnostic plots. Feel free to relegate those to a separate issue if that works better.

[meiertgrootes]

The visualization and the configuration of computational resources are both very nice to have! Histograms of the target residuals the predicted corrections and the error after correction might also be quite informative in judging and demonstrating performance.

[SarahAlidoost]

yes, that's right. I'll merge two notebooks, main and this branch to have both plots.