Bug: Remove reciprocal from monotonic decreasing ops

[lgeiger]

1 / x has an infinite discontinuity at x = 0 (https://www.wolframalpha.com/input/?i=1+%2F+x). Thus it is not monotonic on the entire real line.

The algorithmic optimizer would optimize the case Max(Inv(x)) => Inv(Min(x)). But this is only true for x > 0 or x < 0. E.g.:

x = tf.constant([-1.5, 1., 2.])
tf.reduce_max(tf.reciprocal(x))  # 1.0
tf.reciprocal(tf.reduce_min(x))  # -0.6666667

This PR removes Inv and Reciprocal ops from the list of elementwise monotonic ops.
Fixes #25330 (comment)

/cc @ezhulenev

[ezhulenev]

Whoah, that's scary.

[rmlarsen]

Thanks for catching this.