[Relay] Add tensor rank check for nn.instance_norm

[wzh99]

This PR adds a rank check for input tensor in type inference of nn.instance_norm. I explain the reasons as follows:

First, according to the definition of Instance Normalization, it only normalizes the spatial dimensions. Therefore, the input tensor must be of at least rank 3, and otherwise the operator will not produce meaningful results.

Second, nn.instance_norm with tensor rank less than 2 leads to a problem in the SimplifyInference optimization pass. This pass finds the reduced axes before converting the operator to its lower-level computation definition:

tvm/src/relay/transforms/simplify_inference.cc

Lines 149 to 152 in f15afd2

	Array<Integer> reduced_axes;
	for (int i = 1; i < ndim; ++i) {
	if (i != axis) reduced_axes.push_back(i);
	}

If the rank of the input tensor is less than 3, reduced_axes is empty. According to GetRealAxis, all the dimensions are reduced:

tvm/include/tvm/topi/reduction.h

Lines 67 to 70 in f15afd2

	if (!axis.defined() || axis.size() == 0) {
	for (int i = 0; i < ndim; ++i) {
	real_axis.push_back(i);
	}

This is problematic because we do not actually want batch and axis dimensions to be reduced.

cc @masahi

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: relay _{See #10317 for details}
• Built docs for commit 93f7698 can be found here.

_{Generated by tvm-bot}