NEP-18: CuPy backend requires unimplemented algorithms

For the past few weeks, I've been working on issues related to [NEP-18](https://www.numpy.org/neps/nep-0018-array-function-protocol.html) support for Dask (and ultimately, Dask-GLM as well) to allow CuPy to be used as a backend. I've made some progress, which I'll soon share with more details.

Among the issues I have now is the usage of the following Dask-GLM algorithms with CuPy:

1. `newton`: requires [numpy.linalg.lstsq](https://www.numpy.org/devdocs/reference/generated/numpy.linalg.lstsq.html#numpy.linalg.lstsq)
1. `admm` and `lbfgs`: require [scipy.optimize.fmin_l_bfgs_b](https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.fmin_l_bfgs_b.html)

Both of these functions are not implemented in CuPy. So I'm wondering if anybody knows whether any of these functions have some open and reliable CUDA implementation that could be integrated into CuPy or if we could implement them somewhat easily on top of already supported CuPy functionality?