Segmentation: dice losses are not computed properly when patch does not include every class.

[Zach-ER]

Dice losses should be evaluated as:

(dice_numerator + epsilon) / (dice_denominator + epsilon)

But currently most dice losses in niftynet are missing the epsilon in the numerator. This means that in a situation where classes are not present in the ground truth, the Dice cannot be 1 even if the prediction is perfect.

Concretely, for a two voxel problem,

predicted = tf.constant(
                [[0, 0, 1], [1, 0, 0]],
                dtype=tf.float32, name='predicted')
labels = tf.constant([2, 0], dtype=tf.int32)

The Dice score should be 1.0 (the Dice loss, 1-Score, should be 0.0).