Feature request: Relative force error for active learning.

[Pdeedee]

The absolute force error for selecting structures is not appropriate for some ensembles which covers a broad range of Temperature and Pressure. When the temperature and the pressure is high, the absolute mean of the force is high. The absolute error is in turn higher than the lower temperature force. It makes the threshold changing necessary with the absolute force error when the user is changing from sampling at low temperature to sampling at a higher temperture. With the relative force error, the user can keep the run.in unchanged to sample from high to low temperatures. Moreover, when the same ensemble go through from low temperatures, the relative force error can sample the structures with small force error but with high relative force errors. The relative force error description can be referenced from dpgen:https://docs.deepmodeling.com/projects/deepmd/en/r2/test/model-deviation.html

[brucefan1983]

I believe this will complicate the usage and is not very necessary. An effective way is to set different threshold values for different MD simulations. For example, using a high threshold value in an MD simulation with a high temperature. In practice, it is very difficult to choose a good value for $l$ in the method above.