[tuner] Run baseline benchmarks first with reusable helper function and regression detection

[bangtianliu]

This PR is relevant to address the task in: #783.

1. Run the baseline benchmark at the start to quickly assess whether the tuner is making progress.
2. Run the baseline benchmark both at the beginning and the end of tuning to detect potential performance regressions during the process of tuning (e.g., device overheating).

This PR adds a reusable helper function that takes a list of candidate indices and returns benchmarking results, allowing the same function to be used for baseline and candidate benchmark results.

Furthermore, this PR adds detection and logging for several conditions:

• Uniqueness of baseline results: Verify the uniqueness of baseline results and logging a warning if duplicates are detected.
• Consistency of device IDs across baseline runs: Ensure that the device IDs from two baseline runs match, logging a warning if any discrepancies are found.
• Performance regression detection: Identify performance regressions between two baseline runs and logging a warning if regressions are encountered.
  Also, unit tests are added to show the corner cases.

[Max191]

Overall looks good. Just a few nits about the logging and code reuse. Could you also add a more informative PR description (essentially explain the reasoning for benchmarking baselines twice like in the linked issue).

[Max191]

Looks good, thanks! Just one nit.

[kuhar]

Ideally, these corner cases should be covered by unit tests.

[bangtianliu]

WIP: need to rebase on top of #799, and continue to work.

[kuhar]

Can you edit the PR description and explain what error conditions you considered and how they are handled?