Make initial choice of lambda for Levenberg–Marquardt algorithm configurable

Thanks for this tiny beautiful library! I've been looking for a long time for regressions without unreasonable dependency footprint.

The problem I have is that the initial guess of lambda0 in Levenberg–Marquardt algorithm is too high, like, 1e14, which makes delta too small, immediately triggering relDiff < 1e-10 condition. The workaround is to scale the function so that gradient is gentler, and lambda0 gets below 1e8 or similar:

    maxGrad = maximum $ 1 : map (\(x, _) -> x * log x) xys -- maxGrad is ~1e6
    scale = 1e8 / maxGrad / maxGrad -- scale is ~1e-4

    scaledFits = levenbergMarquardt1
      (\a (x, y) -> (y, mult scale (a * x * log x), mult scale (x * log x)))
      (mult (inv scale) initA)
      xys

    fit = scaleFit scale $ NE.last scaledFits

Could there be a way to specify the initial choice of lambda?

Alternatively, maybe if LM exits on the very first iteration, it might make sense to retry with a smaller lambda automatically.