LP calls instead of MIP calls

[dhendryc]

It can be more useful in some cases to solve the LP in Frank-Wolfe not the MIP (see Network Design Problem).
We currently rely a lot on the fact that the vertices are integer feasible. So this will probably require quite a bit work.

Ideally, we simply create a new Node type and implement the branching and evaluation functions for this (i.e. we will have a new file, e.g. called node_lp.jl).

• Check where we implicitly or explicitly rely on the integrality of vertices.
• Does the solve function rely on that? If yes, how do we best avoid copy-pasting and having duplicate code?
• Does the strong convexity and global/dual tightening work out of the box?
• Adapt strong branching.