feat: add vectorized array_reduce

[osipovartem]

Summary

• add a native higher-order array_reduce(array, initial, merge) function
• evaluate each array position as one Arrow batch and compact inactive rows for uneven arrays
• preserve null-array, null-initial, and terminal null-merge semantics
• add SQLLogic coverage for numeric/string accumulators, coercion, nulls, uneven arrays, and captured columns
• add Criterion coverage for uniform and varying-length arrays

Performance

Measured locally with 8,192 rows and 16-element arrays:

• uniform arrays: ~0.52 ms
• varying lengths: ~1.19 ms

The lambda expression is never evaluated per row. Uniform arrays use a full-batch fast path; uneven arrays evaluate only active rows.

Verification

• cargo +1.94.0 test -p datafusion-sqllogictest --test sqllogictests -- array_reduce
• cargo +1.94.0 clippy -p datafusion-functions-nested --all-targets -- -D warnings
• cargo +1.94.0 bench -p datafusion-functions-nested --bench array_reduce -- --warm-up-time 1 --measurement-time 2 --sample-size 10