fix: Support varying quantile values per group in group_by aggregation

[wtn]

Fixes #20951 and its duplicate #25888.

When using group_by() with a quantile parameter that varies per group (e.g., pl.col.quantile.first()), all groups incorrectly received the same quantile value instead of each group using its own.

Reproduction

df = pl.DataFrame({
    "value": [1, 2, 1, 2],
    "quantile": [0, 0, 1, 1],
})
df.group_by("quantile").agg(pl.col("value").quantile(pl.col("quantile").first()))
# Expected: quantile=0 -> 1.0, quantile=1 -> 2.0
# Actual: both groups returned 1.0

Cause

AggQuantileExpr::evaluate_on_groups() always called get_quantile() which evaluates the quantile expression against the full dataframe, returning a single scalar. This worked for literal quantile values but failed when the quantile expression varied per group (e.g., first() aggregation).

Fix

Added agg_varying_quantile which accepts a slice of quantile values (one per group) and computes quantile per group using the existing aggregation helpers.

polars-core changes:

• Added agg_helper_idx_on_all_with_idx and _agg_helper_slice_with_idx helpers that pass the group index to closures
• Added agg_varying_quantile_generic that iterates over groups with their corresponding quantile values
• Added agg_varying_quantile methods to Float32Chunked, Float64Chunked, integer ChunkedArray, Series, and Column

polars-expr changes:

• AggQuantileExpr::evaluate_on_groups() now detects whether the quantile is uniform (literal/scalar) or varies per group, and dispatches to the appropriate path

[codecov]

Codecov Report

❌ Patch coverage is 93.93939% with 12 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.43%. Comparing base (0a50a14) to head (228ed1d).
⚠️ Report is 188 commits behind head on main.

Files with missing lines	Patch %	Lines
...polars-core/src/frame/group_by/aggregations/mod.rs	93.61%	6 Missing ⚠️
...s-core/src/frame/group_by/aggregations/dispatch.rs	90.56%	5 Missing ⚠️
crates/polars-expr/src/expressions/aggregation.rs	97.56%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##             main   #25606    +/-   ##
========================================
  Coverage   81.43%   81.43%            
========================================
  Files        1801     1801            
  Lines      246750   246930   +180     
  Branches     3081     3081            
========================================
+ Hits       200936   201084   +148     
- Misses      45028    45060    +32     
  Partials      786      786            

☔ 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.

[aparna2198]

@orlp @alexander-beedie @ritchie46 @reswqa @c-peters @MarcoGorelli can we get the review here, we have alot of duplicate issues arising for this one. thanks

[orlp]

I know this was in the original code, but can you make this just a normal unwrap()?

[wtn]

Changed unwrap_unchecked() → unwrap().

[orlp]

Why not match len here for a single-element fast path like below?

[wtn]

Added single-element fast path with match idx.len().

[orlp]

I feel like these are way too many individual tests. You can search around for pytest.mark.parametrize to see how to parametrize the tests. You can verify against a naive Python implementation which groups into lists then manually calls quantile on each list with a specific quantile.

[wtn]

Consolidated into test_group_by_varying_quantile, parametrized over dtype and method, with numpy reference verification.

[wtn]

OK, I've pushed my changes. 🏓