Scope AR MFJ-2dep low-income table rows added in 2024-2025 to those years

[PavelMakarchuk]

Fixes #9024. Backward-leak class (#9008, #9022).

Two late-defined rows in low_income_tax_tables/joint/two_or_more_dependents.yaml carry finite amounts that backfill to earlier instants: (35,000 → $765) from TY2024 resolved at 2021–2023, and (36,001 → $790) from TY2025 resolved at 2021–2024. The published DFA tables end at "Above $32,200" (2021), "Above $33,100" (2022), "Above $34,100" (2023), "Above $35,100" (2024) — verbatim cutoff notes in the table PDFs.

Impact (verified end-to-end pre-fix): TY2021 MFJ + 2 kids at wages $35,050 → ar_income_tax_before_non_refundable_credits_joint $765 via the leaked row, vs $1,092.54 from the regular table (2021 bracket card, verbatim: "$23,600 $39,699 5.0% minus $439.96"). Non-monotonic: a $150 raise cut AR tax by ~$320.

Change: explicit 2021-01-01: .inf threshold+amount on the four late brackets. Post-fix per-year scale behavior: 2021–2023 → inf (falls through to regular table), 2024 → 765/inf, 2025 → published rows.

Tests (expected values from the published tables/cards): 2021 above-cutoff → 1,092.54; 2021 in-table $31,901–$32,000 row → 921 (TaxTables_2021.pdf); 2024 last published row → 765. 14/14 pass.

🤖 Generated with Claude Code

[hua7450]

Verified end to end — approving.

Fix verification (per-year bracket resolution, before vs after): pre-fix, the TY2024 row (35,000 → $765) resolved at 2021–2023 and the TY2025 row (36,001 → $790) resolved at 2021–2024. Post-fix, each year ends exactly at its published cutoff — 2021 → "Above $32,200", 2022 → $33,100, 2023 → $34,100, 2024 → $35,100 — and the 2025 column is bit-identical to main. Confirmed the published rows verbatim against the 2021 instructions, 2024 instructions, and TaxTables_FI_2025 PDFs.

Completeness: scanned all eight AR low-income table files for the same backfill class. The other seven files' late-added brackets all backfill .inf amounts (benign — redundant cutoffs above each year's real cutoff). This file was the only one with harmful finite-amount backfills, so the fix is complete for the leak class.

Tests: variable is entity = Person, so the array outputs are correct; all three expected values re-derived independently (35,050 − 4,400 = 30,650 taxable; 0.05 × 30,650 − 439.96 = 1,092.54 ✓; $921 row ✓; $765 row ✓).

Follow-up (pre-existing, filed separately): the 2021–2024 columns key rows at X_X00 while the published tables run $X,X01–$X,X00 inclusive (the 2025 column is correctly keyed at X_X01). At exact-$100 incomes the wrong row applies — e.g., 2024 income of exactly $35,000 returns $765 instead of $754, and exactly $35,100 falls to the regular table instead of $765. Out of scope here since this PR only adds the .inf scoping.