[tuner] Add candidate generator script

[RattataKing]

The plan is to move the tuner from the sdxl-scripts repo to the sharktank repo.
This PR is the first step, moving candidate_gen.py (the candidate generator for tuning) along with its CI and unit tests.
More changes will come to fully decouple it.

New changes in candidate_gen.py:

• Edited comments
• Added a registry class for dispatch kind
• Moved dispatch kind specific functions from global to their belonged classes
• Changed how it walks op and when it decides which functions to use for candidate generation.
• Checked if input mlir translation info contains LLVMGPUVectorDistribute

candidate_gen_test.py:

• No new tests added, updated the existing function to match changes in candidate_gen.py

[kuhar]

Please add a description to this PR and say what the plan is wrt landing the tuner.

Also please include the unit tests together with the code.

[ScottTodd]

A few comments from an initial scan. Definitely want to resolve the github workflow / requirement file topics before merge here. Could defer the script refactoring.

[ScottTodd]

A README.md in sharktank/tools/tuner/ would be helpful too

[kuhar]

LGTM, thanks for all the cleanups

[kuhar]

The new location looks good to me