fix: Reserve fixed-size-list child capacity

[0guban0v]

Closes #28872

More measurements are underway, but PR is ready for review. I realized it has a bigger blast radius than just arr.dot.
Added comment with broader impact test.

Benchmark used arr.dot streaming broadcast, width 768, 16,776,960 products, and validated native/composed equality.

Revisions: main a6782fa065, PR 56ddea3e98

100 × (composed heap allocation − native heap allocation) / (rows × width × sizeof(inner_dtype))

+100% → native avoids heap allocation equal to one product buffer
   0% → equal heap allocation
 −33% → native allocates 33% of one product buffer more

Early numbers:

Inner dtype   1t before   1t after   8t before   8t after  16t before   16t after
Float32       -33.3280%   +0.0049%   -33.2688%   +0.0647%   -33.2005%   +0.1342%
Float64       -33.3282%   +0.0050%   -33.2827%   +0.0506%   -33.2300%   +0.1023%
Int8          -33.3262%   +0.0091%   -33.1841%   +0.1491%   -33.0234%   +0.3106%
UInt8         -33.3276%   +0.0078%   -33.1845%   +0.1493%   -33.0188%   +0.3119%
Int16         -33.3266%   +0.0062%   -33.2402%   +0.0930%   -33.1385%   +0.1920%
UInt16        -33.3280%   +0.0062%   -33.2401%   +0.0930%   -33.1395%   +0.1914%
Int32         -33.3281%   +0.0052%   -33.2686%   +0.0646%   -33.2013%   +0.1324%
UInt32        -33.3284%   +0.0057%   -33.2685%   +0.0647%   -33.2022%   +0.1322%
Int64         -33.3284%   +0.0050%   -33.2828%   +0.0507%   -33.2302%   +0.1030%
UInt64        -33.3280%   +0.0051%   -33.2828%   +0.0506%   -33.2305%   +0.1031%
Int128        -33.3286%   +0.0047%   -33.2897%   +0.0437%   -33.2453%   +0.0883%
UInt128       -33.3286%   +0.0048%   -33.2897%   +0.0436%   -33.2452%   +0.0881%

Post-fix, native arr.dot allocates less heap than composed, and that saving grows with thread count.

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 81.33%. Comparing base (a6782fa) to head (c25a47c).
⚠️ Report is 99 commits behind head on main.

Files with missing lines	Patch %	Lines
.../polars-arrow/src/array/fixed_size_list/mutable.rs	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #28980      +/-   ##
==========================================
- Coverage   81.58%   81.33%   -0.26%     
==========================================
  Files        1890     1886       -4     
  Lines      268762   267093    -1669     
  Branches     3264     3061     -203     
==========================================
- Hits       219280   217246    -2034     
- Misses      48637    49069     +432     
+ Partials      845      778      -67     

☔ 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.

[0guban0v]

To test broader impact, I isolated generic construction path DataFrameBuilder::reserve → SeriesBuilder::reserve → FixedSizeListArrayBuilder::reserve

I directly compare reserve(additional) with reserve(additional * size). Each case processes 4,194,048 Int64 child values at widths 1, 8, 64, 768. Both modes reserve all outer rows first: single extend copies input once, while repeated subslice_extend copies approx 65,536 child values per call.

Revisions: main 73dd5bea65, PR c25a47ca73.

Arrows in table mean main → PR.
100% means final child capacity equals required payload.

Mode	Width	Growth reallocations	Final child capacity / required payload
single extend	1	0 → 0	100.000% → 100.000%
single extend	8	1 → 0	100.000% → 100.000%
single extend	64	1 → 0	100.000% → 100.000%
single extend	768	1 → 0	100.000% → 100.000%
repeated subslice_extend	1	0 → 0	100.000% → 100.000%
repeated subslice_extend	8	3 → 0	100.000% → 100.000%
repeated subslice_extend	64	6 → 0	100.000% → 100.000%
repeated subslice_extend	768	8 → 0	199.231% → 100.000%

Corrected reservation eliminates every observed child-buffer growth reallocation for widths above 1. Most cases reach same capacity without intermediate growth, but width 768 also reduces final capacity from 199.231% to 100% of required payload.