perf: Speed up dt.offset_by 2x for constant durations

[MarcoGorelli]

closes #16722 (timing becomes in line with the alternative method presented - I couldn't reproduce the 10x difference reported)

timing results available here https://www.kaggle.com/code/marcogorelli/polars-timing?scriptVersionId=181508652

---

I think this is fine but I didn't sleep too well so I'll check over it again tomorrow before marking as 'ready for review'

---

The gist of the change is: in the simple case, keep things simple and use apply_values

[mcrumiller]

I think this makes the same duration_*s() call for every element, right? You can pull the offset out front and then apply the value, as in:

let offset = match datetime.time_unit() {
    TimeUnit::Milliseconds => offset.duration_ms(),
    TimeUnit::Microseconds => offset.duration_us(),
    TimeUnit::Nanoseconds => offset.duration_ns(),
};
Ok(datetime.0.apply_values(|v| v + offset)

[MarcoGorelli]

thanks @mcrumiller!

[codspeed-hq]

CodSpeed Performance Report

Merging #16728 will not alter performance

_{Comparing MarcoGorelli:faster-offset-by (42777b5) with main (5d0e339)}

Summary

✅ 37 untouched benchmarks

[codecov]

Codecov Report

Attention: Patch coverage is 96.66667% with 1 line in your changes missing coverage. Please review.

Project coverage is 81.45%. Comparing base (79afe75) to head (42777b5).
Report is 11 commits behind head on main.

Files	Patch %	Lines
...ates/polars-plan/src/dsl/function_expr/temporal.rs	96.66%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main   #16728   +/-   ##
=======================================
  Coverage   81.45%   81.45%           
=======================================
  Files        1413     1413           
  Lines      186096   186057   -39     
  Branches     2776     2756   -20     
=======================================
- Hits       151585   151557   -28     
- Misses      33991    33997    +6     
+ Partials      520      503   -17     

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

[ritchie46]

Just notice this now, but ideally the implementation should be in polars-ops or polars-time and then we only dispatch to those functions here. I'd like polars-plan to only be about the query planning/resolving.

[MarcoGorelli]

yup - ok if I do that separately?

[ritchie46]

To reduce binary bloat, I think we can make one code branch of this that does the apply_values.

if offset.negative() {
   duration = - duration;
}

Ok(datetime.0.apply_values(|v| v + duration))

I think we can also use our already compiled kernels now. Which might also have more SIMD acceleration. If not, it at least saves compiler bloat.

// apply to inner and then you have to reconstruct the datetime.
datetime.0.clone().wrapping_add_scalar(duration);

[ritchie46]

Great hammering on those temporal performance! 💪 Left some comments.