Add a custom easy-to-understand "first approach" optimization

[mtewes]

I have in mind something brute force, that respects the idea that our networks are small, start with something already reasonable, and that "sparse" changes can have a huge effect. Example: loop through all params, try out -1, -0.5, +0.5, +1, and then perform a single step changing just this parameter.

[mtewes]

I started working on this, but will take care of #18 at the same time. The purpose of all this is to see why the current bfgs gets stuck in "obviously bad" solutions.

[mtewes]

Now restructuring the training, to support the choice of algorithm (or several, in fact) at a high level (megalut wrapper).

[mtewes]

@kuntzer I've made PRs for both tenbilac and megalut. They go together (you need both), but I guess that's ok given our current stage of experimentation.

[kuntzer]

Okay, I'm looking at it now

[mtewes]

Merged. It's still a mess to pass arguments through the megalut tenbilacwrapper, but at least this makes it cleaner to write e.g. your own hard-code optimizer (for a specific task) and using it.