find way to generate randomized training data

[cburgdorf]

The idea goes roughly like this:

We write a generator that generates subsequent moves in our grid. Our stones are indicated as 1, the other player is -1 and an empty field is just 0.

1. The generator starts with an empty field
2. It puts a stone somewhere (the generator knows stones fall down top to bottom)
3. It takes the previous grid and puts another stone somewhere. It's always a random action between 1 and 6 to pick a slot. The stone will then simply fall down (no randomness there)
4. This goes on until we hit a connect four or the grid is full

Out of all the data we just take the ones that lead to a win and are within a threshold of move counts. After all we aren't that much interested in bad games (I think).

This should build up a nice collection of training data.

@Lofavreel Does that sound reasonable to you?