Fix subsecond precision in array timestamp offsets

[farzaandalwai]

Array step counts currently lose subsecond non-Tick offsets when pandas timestamps use a coarser resolution. For example, a 500 ms DateOffset applied to datetime64[s] values is retained with scalar n but silently truncated with array n. This addresses the offset_times portion of #304; the separate time_ranges behavior remains outside this change.

The fix concatenates the computed runs without an empty input-resolution prefix or a final narrowing cast. Arrow-backed inputs retain their storage type with the computed timestamp resolution. It also adds a focused README example and an offset_times docstring.

Validation:

• The new regression cases initially produced 14 failures and 14 passes without the fix.
• The full suite passed locally on Python 3.10.21: 1,829 passed, eight skipped, 90.56% coverage. Spark required Java 21 and its worker Python to match the driver.
• pre-commit checks for utilsforecast/*.py, the new test file and README passed: Ruff lint, Ruff format and mypy.
• The tests cover mixed positive/negative/zero counts, second/millisecond/microsecond input resolutions, Series/Index metadata, timezone-aware values, Arrow storage and empty inputs.

AI assistance: OpenAI Codex assisted with the implementation, tests, documentation and this description. The validation results above were executed locally. No changes are proposed to time_ranges or the existing Arrow-date pull request #303.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[José Morales (jmoralez)]

Superseded by #308