three_nn gpu implementation.

[azaganidis]

Redefinition of three_nn by tensorflow operations instead of c++ on cpu. Significantly faster.

[hi-zhengcheng]

If it returns directly from tf.nn.top_k(-dist, k=3), I think it's wrong. Pay attention to the minus sign you added before dist. So it should be modified as:

values, indices = tf.nn.top_k(-dist, k=3)
return -values, indices

Nice implementation.