feat: report absolute eval time in nanoseconds instead of relative to reference

[ArpanC6]

Closes #1374 (partially)

What was the problem?

The benchmark reported t(eval) as a ratio relative to an arbitrary
reference function (simple_assume_observe_non_model), which made
results hard to interpret and added unnecessary noise.

What did I fix?

• Removed the reference time baseline
• Now reporting t(eval) in absolute nanoseconds
• t(grad)/t(eval) ratio is kept as-is
• Updated column headers accordingly

Note

Introducing Stan as an external baseline is a separate more involved
task left for a future PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.35%. Comparing base (10a3651) to head (03079d5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1383   +/-   ##
=======================================
  Coverage   82.35%   82.35%           
=======================================
  Files          50       50           
  Lines        3531     3531           
=======================================
  Hits         2908     2908           
  Misses        623      623           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ArpanC6]

Hlw @yebai, this PR partially addresses #1374.

I've gone through the issue carefully and made the following changes

• Removed the arbitrary reference time baseline (simple_assume_observe_non_model)
• t(eval) is now reported in absolute nanoseconds, which is more meaningful and interpretable
• t(grad)/t(eval) ratio is kept as-is since it was already useful
• Updated column headers accordingly

I also noted @penelopeysm's point that introducing an external baseline doesn't simplify comparative benchmarks you still need two checkouts regardless. So I've kept this PR focused on the absolute time reporting for now.

Note: The Benchmarking / combine results check is failing but this appears to be a pre existing CI issue unrelated to this PR.

Introducing Stan as an external baseline would be a separate more involved task happy to work on that as a follow up PR.