Change default learning rate in Descent Optimizer

[TheFibonacciEffect]

Sets default lerning rate of Descent() optimizer to a lower value, to make models more likely to converge.

Closes #2623

PR Checklist

• Tests are added
• Entry in NEWS.md
• Documentation, if applicable

Reasoning

It is more important that a solver converges rather than that the convergence is done quickly.

I will add the code in the Issue as a test case later, before merging.