fix: handle nested RANGE CURRENT ROW peers

[sunchao]

Which issue does this PR close?

Related to #24937. This fixes the RANGE-window case; general MIN/MAX scalar comparisons remain separate.

Rationale for this change

A RANGE window ordered by a list containing lists or structs can plan successfully and then fail with Uncomparable values. For example:

SELECT count(*) OVER (ORDER BY l)
FROM (VALUES ([[1]]), ([[2]])) t(l);

CURRENT ROW boundaries only need to find equal ordering keys in the sorted input.

What changes are included in this PR?

Use Arrow's comparator to find nested CURRENT ROW peer boundaries. Reuse the comparator while the ordering arrays are unchanged and release it when buffered rows are pruned. Preserve WindowFrameStateRange's existing unwind-safety traits.

SQL type admission, scalar-key handling, and finite-offset calculations are unchanged.

What is the testing strategy for this PR?

Update the two existing SQL regressions from expected errors to successful counts. Physical-operator tests cover nested lists/structs, nulls, tied ordering keys across batches, frame boundaries, and completed-peer emission before EOF in sorted mode. A small state test covers comparator reuse, pruning, and the public unwind-safety trait guarantees.

Local validation on this patch:

• Full extended workspace suite: 12,421 Rust tests passed, eight ignored, and all 527 SQL-logic files completed.
• All-target/all-feature Clippy with warnings denied passed. The downstream catch_unwind compatibility reproducer also compiles for both borrowed and owned window state.
• Formatting, license/typo checks, generated documentation checks, local link checks, and the security audit passed. API and website documentation builds also passed with warnings treated as errors.
• Full lint could not complete locally: the package mirror does not offer cargo 0.100.0, pinned by the existing dev/depcheck tool. This blocks the circular-dependency check and that tool's metadata inspection by cargo-machete; cargo-machete otherwise reported no unused dependencies. This patch changes no dependency manifests or lockfiles.
• The physical regression and both SQL regressions fail on the unpatched target base with the original comparison error.
• All six existing aggregate-window benchmark smoke cases pass on both base and candidate. No throughput measurements are claimed.

Are there any user-facing changes?

Already-admitted SQL RANGE windows over nested list keys return results instead of an internal comparison error. The same boundary fix applies to windows constructed through the physical API. No new public API or SQL-admitted types are introduced.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 87.61905% with 13 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.72%. Comparing base (c0e872f) to head (f0d7123).
⚠️ Report is 20 commits behind head on main.

Files with missing lines	Patch %	Lines
datafusion/expr/src/window_state.rs	87.25%	6 Missing and 7 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #26067      +/-   ##
==========================================
+ Coverage   82.69%   82.72%   +0.02%     
==========================================
  Files        1147     1147              
  Lines      447348   449092    +1744     
  Branches   447348   449092    +1744     
==========================================
+ Hits       369936   371509    +1573     
+ Misses      54999    54932      -67     
- Partials    22413    22651     +238     

☔ View full report in Codecov by Harness.
📢 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.