Skip adding control_deps if nodes are the same

[reminisce]

The following code would fail before this PR due to introducing a cycle into the gradient graph. See this issue for analysis: #8029. This PR also fixes this issue of the same root cause: #7956.

data = mx.symbol.Variable('data')
res = data + data + data + data + data + data + data + data
res.simple_bind(ctx=mx.cpu(), data=(1,))

@piiswrong @tqchen

[tqchen]

LGTM, can you backport this to nnvm? Thanks

[piiswrong]

@tqchen Do you remember why the inplace_sum_cap was there? Can we remove it?

[reminisce]

@tqchen Could you specify where in nnvm I should backport this fix to?

[tqchen]

@reminisce nevermind, this logic is only in mxnet

[tqchen]

inplace sum is used to rewrite normal sum to a chain of inplace sum, which saves more space. Since if we do an explicit sum for weight gradient of RNN, we cost O(timestep) memory