Update learning rate schedule in training/default.yaml

[jswijnands]

This updates the default values for the learning rate schedule. Specifically, the local learning rate is increased to 8x the original learning rate. In addition, the number of iterations is reduced from 300k to 150k. This has led to similar accuracy in stretched grid experiments using the graph transformer architecture, while requiring only 50% of the GPU compute. Optimised default values could save GPU resources across the Pilot Project.

Suggested review questions:

• Are these settings also an enhancement for global / LAM models? (tested only for stretched grid)
• Are these settings also an enhancement for GNN / transformer models? (tested only for graph transformer architecture)

[FussyDuck]

All committers have signed the CLA.

[JesperDramsch]

Hi @jswijnands, thank you for the contribution.

We'll look into benchmarking this on the non-stretched grid models and evaluate if we should maybe make this a special config, as I believe the stretched-grid may generally benefit from different training defaults.