[Relay][VM] Add ReshapeTensor instruction in the VM to replace the reshape op

[icemelon]

@jroesch @zhiics @mbrookhart Could you help review?

[zhiics]

lgtm

[zhiics]

Please check the error. We have gpu context in the check_result.

[mbrookhart]

LGTM

[mbrookhart]

The GPU error seems to come from the check I added to fail early on GPU with dynamic models. I believe dynamic shapes wont work on cuda until we get heterogenous execution working for shape funcs, per our discussion in https://discuss.tvm.ai/t/issue-with-vm-execution-when-testing-gpu-immediately-after-cpu/7010/7

@zhiics, any update on that work? Or should we limit this test to CPU?

[zhiics]

@mbrookhart Yeah, I meant there is gpu context in the test. That's why the ci fails. We can probably just disable check_result for this test ATM. Heterogeneous execution needs this PR.

[zhiics]

Thanks @icemelon9 @mbrookhart

[MarisaKirisame]

@icemelon9 is_reshape_only is incorrect. If a function return a constant it will wrap a reshape on it. Please fix it in a principle way (e.g. not by checking that the whole function has at least one reshape, as this will not work on reshape of constant).