Look into precision of unitballs parameterization

[ludwigschubert]

tests/optvis/param/test_unit_balls.py occasionally fails in the ~shape=(2, 5, 5), precondition=True~ all cases for L1 balls with tf.reduce_max values observed as high as ~1.08~ 1.3.

[ludwigschubert]

@colah On re-observation, being precise to "only" 1e-7 actually feels in line with machine precision for float32, no?

I still feel like in the L1 case where we essentially divide a number by itself we should get a precise 1.0, but that may just be Tensorflow graph weirdness.

Entirely unexplained are the results we see on Travis. (I've seen ~1.3 by now)