Count elapsed time across DST in timesince

[pralav-25]

timesince subtracts the input from datetime.now(input.tzinfo). When both share the same changing timezone, Python ignores UTC-offset changes and fold: the two occurrences of 2026-11-01 01:30 in New York are one hour apart but currently produce just now. Across the spring transition, one elapsed hour can produce 2 hours ago.

Convert the current aware time to UTC before subtraction so elapsed time includes offset changes. Leave the input in its original timezone to preserve valid datetime.min/max values whose equivalent UTC dates would overflow. Naive datetime, timedelta and date-error behavior stay the same. Add past/future fall-back and spring-forward regressions, boundary coverage, and a short guide clarification.

Validation:

• Four DST regressions fail against the original implementation; the two boundary cases already pass and remain covered.
• Full tox passes on CPython 3.10–3.14 and PyPy 3.10–3.11: 348 tests per interpreter and 100% statement and branch coverage.
• Ruff check/format, mypy, basedpyright, pyrefly, codespell, Sphinx with warnings treated as errors, isolated wheel/sdist builds and Twine checks pass.
• Independent review verified 150 elapsed-time cases, including New York, Berlin, Lord Howe half-hour transitions and Apia's skipped day, plus six extreme date/offset cases and existing naive/date/custom-timezone behavior.

Prepared with OpenAI Codex assistance, including implementation, tests and independent review. Current and historical issues and pull requests, including the aware-datetime support in #54, were checked for duplicates.