v1.0.0b27

• Make model building fully deterministic (#821).
• Implement retry mechanism for 429 web error GLOPOP-SG data adapter. Zenodo rate limited making too many range requests. If we get rate limited now, we now go to sleep for a bit and try again later.
• Further reduce memory use of reporter by using numpy views for identical time series data.
• In a recent change, discharges were explictly set to nan in reservoirs (good). This led to a case where rivers had 0 river width in the hydrodynamic model, which led to raised errors. Now, we fix this by looking further downstream of reservoirs to find a valid discharge for estimating river widths. This is also better, because perviously the default alpha for river width was used.
• Add geb tool merge command to merge outputs from multiple GEB cluster sub-models into a single merged model directory that can be evaluated directly with geb evaluate.
• Add plot_skill_score_maps() to plot skill scores per station on a satellite basemap for each metric.
• Update plot_skill_score_boxplots() with violin + boxplot panels for KGE, NSE, R², RMSE and RRMSE, including support for overlaying external reference model (e.g. LISFLOOD) results.
• Add R², RMSE and RRMSE to discharge evaluation metrics at hourly, daily and monthly frequencies.
• Stations with fewer than 5 years of paired observations are skipped in evaluate_discharge.
• Lakes and reservoirs are now shown as dot markers in the interactive discharge evaluation map.
• Include height above nearest drainage (HAND) for later use.