[Numpy] Unique has not registered the gradient

[sxjscience]

import mxnet as mx
from mxnet.gluon import nn, HybridBlock
mx.npx.set_np()

class Foo(HybridBlock):
    def hybrid_forward(self, F, x):
        y = F.np.unique(x)
        return y

foo = Foo()
foo.hybridize()
out = foo(mx.np.ones((10,)))
print(out.asnumpy())

Error message:

MXNetError: Traceback (most recent call last):
  File "../src/nnvm/gradient.cc", line 213
MXNetError: Operator _npi_unique is non-differentiable because it didn't register FGradient attribute.

[haojin2]

Root case is that the gradient for unique is not well-defined, so its gradient function was not registered at all in the first place. As per offline discussion with @sxjscience, we decide to add a zero grad node for the gradient.