OOM for large datasets

[havardhhaugen]

On lumi I can run an n320+2.5km checkpoint on dev-g if I use a small dataset (1 date), but it runs out of memory if I change this to a year. This should not be the case since data is read batch-wise and predictions are stored on the fly.

It runs out of memory at some point during the first prediction, so its probably not a memory leak.