Enable cpp code path for Categorify ops

[rjzamora]

Rolls back one of the changes in #354 to improve inference performance of Categorify.

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/systems/review/pr-389

[oliverholworthy]

the docs related errors are now resolved.

There's two remaining things for this MR:

• fix the lint errors (reformat with black pre-commit run --all-files black) and the codespell error (typo in everything)
• merge the NVTabular PR Fix Categorify inference and testing NVTabular#1874
  • currently blocked by some unrelated errors with docs (which will be resolved by Update Docs Theme to use sphinx-book-theme NVTabular#1878 )

[rjzamora]

There's two remaining things for this MR: ...

Thanks so much @oliverholworthy ! Sorry, I missed this earlier.

[oliverholworthy]

Following the merge of NVIDIA-Merlin/NVTabular#1874 we're now getting a new error in the tests

RuntimeError: Error: <class 'ValueError'> - Unknown column for CategorifyTransform x__values

[rjzamora]

I haven't had the time to understand the test failure well enough to work through a long-term fix. However, my last commit roles back the decision to use the cpp code path for "Categorify" by default. Instead, when the WorkflowRunner is initialized, the value of the NVT_CPP_OPS environment variable will dictate which NVT ops the cpp code path will be used for.

For example, you can enable the cpp version of Categorify by deploying a triton model as follows:

NVT_CPP_OPS="Categorify" tritonserver --model-repository=/my-model/

[jperez999]

This is failing right now on some package mismatches right now. A solution is to update the tritonclient package to 2.30.0. Working on that solution now. Then we should be able to pick up the newest version of CI container to fix the triton errors.