Add CrossEntropyLoss

[MoritzWillmann]

Current State

This PR adds a new loss class CrossEntropyLoss implementing the multi-class log loss (aka. cross entropy loss)

The value is calculated as

$$L = -\frac{1}{N} \sum_{i=1}^N \sum_{j=1}^K y_{ij} \cdot \log\left(f(x_i)_j\right)$$

If the task at hand is binary, a "class" is added with prediction $1-f(x)$ and label $1-y$.

The gradient is thus calculated as

$$\frac{\partial}{\partial \theta} L = -\frac{1}{N}\sum_{i=1}^N \sum_{j=1}^K \frac{y_{ij}}{f(x_i)_j} \frac{\partial f}{\partial \theta}$$

If the task at hand is binary, a "class" is added with prediction $f(x)-1$ and label $1-y$.

I test this functionality in tests/qnn/test_loss.py.

This PR solves #112

[MoritzWillmann]

Update

I figured out that if if eps is sufficiently large, training works. The idea of eps being dependent on the learning task and model at hand, i.e. making it somewhat of a hyperparameter, seems very strange to me. Maybe we can gain some more insights on this.

[MoritzWillmann]

Not sure, why this was closed

[rothma]

@MoritzWillmann in my tests it works now (at least for binary)

[MoritzWillmann]

I finally was able to fix it. LGTM 🎉