Add ResMLP test

[slyubomirsky]

This PR adds an end-to-end run of ResMLP (from PyTorch import to running) and compares the TVM result against the PyTorch result.

There is an additional rewrite here of linear layers. The TVM PT importer produces linear layers that look like this

reshape(nn.dense(*, *), (1, 256, 512)) + *

whereas the FlexNLP codegen matches patterns of the form nn.bias_add(nn.dense(*, *), *).

Fortunately, the above is equivalent to reshape(nn.bias_add(nn.dense(*, *), *), (1, 256, 512)), which does match the pattern. The rewrite is currently implemented manually, but this would be a useful test case for flexible matching in the future.

Please review @AD1024 -- you will have to add in the calls to our custom codegens.

[AD1024]

Thank you for implementing this. I tried to annotate the module with flexnlp and ilavta codegen, but it seems that the rewrite does not work properly (linear pattern is still not matching for flexnlp). Maybe I missed anything?

Also, Conv1D can be offloaded to ILA simulator but the size is too large to terminate.

[slyubomirsky]

I think you're right, the rewrite isn't firing. I'll dig into it some more

For Conv1D, we could try to use a smaller model size

[slyubomirsky]

It was a simple bug (I was not checking for the operator's name when I should have)

[slyubomirsky]

Note: For now the rewrite I have written here will be correct in the ResMLP model provided. There need to be more conditions on the dimensions for it to work in general. That might also be an interesting use case for flexible matching, since the rewrite system would probably have to reason about the shapes of all the terms.