[PERF] simplify conditional_join result materialization (blocked by #1633)

[samukweku]

Dependency

Blocked by #1633. Do not merge this PR until #1633 has merged. This PR is intentionally based on issue-1627-anti-join so the benchmark and refactor cover left_anti and right_anti. After #1633 merges, retarget/rebase this branch onto dev and mark this PR ready for review.

Closes #1632.

Summary

• replace branch-specific result assembly with synchronized left/right row indexers
• use pandas dtype-aware take only when a side contains missing positions, retaining a direct-indexing fast path otherwise
• preserve row order, indicator categories, NumPy promotion, and extension-array dtypes
• remove the copied pandas private fill-array helper
• add mixed-dtype compatibility tests

ELI5

A join first finds matching row numbers, then builds a new table from them. The old code rebuilt that table differently for every join type. This change creates one row-number plan per side; -1 means that side has no row and pandas supplies the correct missing value.

Performance

Apple Silicon, Python 3.14.5, pandas 3.0.3, 50,000 rows, sparse duplicate-key matches, five repetitions. Timing and tracemalloc memory passes ran separately using a one-off local harness that is not included in this PR.

Join	9 cols	32 cols
left	1.04x	1.22x
right	1.21x	1.23x
outer	1.00x	1.10x
left_anti	1.18x	1.05x
right_anti	1.16x	1.16x

Nine of ten sampled cases improved; narrow outer was effectively flat (4.72 ms to 4.74 ms). Peak traced allocation fell in eight cases, was effectively flat for wide outer/right_anti, and rose 3.7% for narrow outer because one combined integer row indexer remains live. The broader 48-case join-type/density/width run found no material runtime regression. End-to-end timing is matching-dominated and stayed within the unchanged inner-path noise band.

Verification

• project suite: 1,346 passed, 5 skipped, 48 xfailed, 17 xpassed
• conditional_join focused suite: 253 passed, 1 skipped
• mixed-dtype materialization regression tests: 12 passed
• configured pre-commit hooks: passed for all changed files
• standalone markdownlint: passed for changed Markdown files
• adversarial comparison against [ENH] Add anti joins to conditional_join #1633: exact parity for numeric/nullable/category/string/datetime/tz/timedelta plus Period, Interval, SparseArray, and empty-input probes