Maybe refine lambda sequence for alpha close to zero

[erikcs]

See #38. Also, maybe the glmnet type threshold of 1e-3 is a bit low in the covariate balancing setting.