[codex] fix historical timezone endOf regression

[susiwen8]

Summary

• fix endOf() boundary calculations to derive the last instant from the next unit boundary instead of constructing local 23:59:59.999
• add a regression test for issue daysInMonth() returns 1 when timezone is Asia/Kuala_Lumpur #3003 covering the historical midnight offset shift in Asia/Singapore / Asia/Kuala_Lumpur
• run the regression under TZ=Asia/Singapore in the test script so the bug stays covered in CI

Root Cause

daysInMonth() relies on endOf('month').date(). For 1981-12-01 in Singapore and Kuala Lumpur, constructing the local month end as 1981-12-31 23:59:59.999 crosses the midnight offset jump to UTC+08:00, so the native Date normalizes to 1982-01-01 00:29:59 and the day becomes 1.

This patch follows Moment's approach and computes endOf(...) as startOf(nextUnit) - 1ms, which preserves the correct local last day even when midnight itself has a historical offset change.

Validation

• TZ=Asia/Singapore npm run test-issue3003
• TZ=Asia/Kuala_Lumpur npx jest test/issues/issue3003.test.js --runInBand --coverage=false
• npx jest test/display.test.js test/manipulate.test.js --runInBand --coverage=false
• npx jest test/plugin/badMutable.test.js --runInBand --coverage=false
• npx jest test/plugin/timezone.test.js --runInBand --coverage=false -t "startOf and endOf"
• TZ=Asia/Singapore npm run test-tz
• npx eslint src/index.js test/issues/issue3003.test.js
• npm test runs with 94/94 suites and 776/776 tests passing, but currently exits non-zero because the repository's global coverage threshold still reports lines 99.95% < 100%