[Relay][OP] Add reverse_reshape

[icemelon]

Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.

[junrushao]

It somehow makes the API over-complicated compared with numpy's, but I do believe it is okay to be a super set of numpy, so it looks good to me

[jroesch]

Not against this change, but curious why do we need this?

[icemelon]

Some MXNet models use "reverse" to reshape tensors. It seems unlikely to support it without this argument since it requires alignment from the right to infer shape.

[tqchen]

While this makes things easy for MXNet in certain cases, I feel we should not introduce it in the normal reshape, or introduce an experimental operator(_contrib_reverse_reshape, level=10) to support such a case.

It would be great to keep the core IR simple and elegant.

[icemelon]

@tqchen Sure, I can move this to an experimental operator.

[icemelon]

@tqchen Now implements reverse_reshape as a level-10 op. In C++, I re-use the code of reshape infer_type to reduce the code duplication. But Python and C++ sees reshape and reverse_reshape as two different ops.
@vinx13 @junrushao1994 Could you review this again? Thanks.

[tqchen]

@icemelon9 please rebase against the master and you can merge the PR

[icemelon]

@vinx13 @junrushao1994 @jroesch @tqchen Thanks for the review.
@tqchen Rebased to master now. Could you approve the PR? It requires 1 approving review from committers.

[yzhliu]

Thanks @icemelon9 @junrushao1994 @jroesch @vinx13 @tqchen now it is merged.