[ML] Include memory required for persistence in data frame analytics memory estimates

The memory estimates used in data frame analytics should include an allowance for the extra memory required during persistence and restoration. Otherwise during persistence or restoration on a maxed-out node there is a risk that some process (not necessarily the job doing persistence or restoration) gets killed because there isn't enough memory to accommodate both the model and the extra data structures used during persistence and restoration.

It would be reasonable that this is just a simple multiplier of the estimated model size. If the main extra memory usage is in compressed form then maybe 10% would be enough, or maybe 10% plus a capped amount for a single uncompressed chunk, for example, `reported memory estimate = raw memory estimate * 1.1 + min(16MB, raw memory estimate)`.

Commencez par localiser le code d’estimation de la mémoire de data frame analytics ainsi que les chemins de persistance et de restauration. Comparez l’estimation brute du modèle avec la mémoire transitoire supplémentaire utilisée pendant ces opérations. Le travail est terminé lorsque l’estimation rapportée inclut une marge convenue et que les tests couvrent la persistance ou la restauration lorsqu’un nœud est proche de sa limite de mémoire.

Rédigé par le modèle d'indexation à partir du texte de l'issue.