Better initial L-BFGS stepsize as default leads to slow convergence in many cases.

[CompRhys]

See #553 -- this mirrors prior issues highlighted to me where the out of the box experience of the torch-sim L-BFGS is not as fast to converge as ASE.