Move out Lagrangian.jl part

[lkapelevich]

The part of the library that solves the Lagrangian dual could be stand-alone and could be made more generic, perhaps for other cut types too.

[odow]

The trend with MOI has been to build bigger repos since it is easier to maintain in comparison to lots of fiddly dependencies.

[bfpc]

I rather have Lagrangian.jl be independent of SDDiP, especially since its functionality is very orthogonal to the "binarization" part, or changing the backwards pass of SDDP. Of course, now there is the issue of moving this to JuMP-0.19 / MOI...

Also, a Lagrangian-MOI is a good step towards augmented-Lagrangian MOI ;-)

[lkapelevich]

The Lagrangian.jl stuff is indeed independent. For moving in with the new SDDP.jl I was thinking it would be best to carry over just one method specialized for SDDiP. But a separate package for Lagrangian relaxations would be fun :-).