Preserve sub-day seconds for precisedelta month and year minimum units

[PDGGK]

Changes proposed in this pull request:

• Preserve sub-day elapsed seconds when rounding precisedelta to a minimum unit of months or years.
• Divide the whole elapsed-second count directly by the scaled month/year divisor, then consume the lower components. This avoids both dropped seconds and intermediate double rounding.
• Add regressions for fractional units, short durations, and rounding midpoints. Keep the existing 30.5-day month, 365-day year, negative-duration behavior, and documented microsecond-precision limitation.

Before and after

from datetime import timedelta
from humanize import precisedelta

precisedelta(timedelta(days=15, hours=6), minimum_unit="months", format="%.2f")
# Before: '0.49 months'
# After:  '0.50 months'

precisedelta(timedelta(days=30, hours=12), minimum_unit="months", format="%.2f")
# Before: '0.98 months'
# After:  '1 month'

precisedelta(timedelta(days=182, hours=12), minimum_unit="years", format="%.4f")
# Before: '0.4986 years'
# After:  '0.5000 years'

precisedelta(timedelta(hours=12), minimum_unit="years", format="%.2f")
# Before: '0 years and 43200000000 microseconds'
# After:  '0 years'

Validation

Linux, CPython 3.12.14:

• Final full suite and source doctests: 883 passed, 2 translation-update tests skipped because GNU xgettext is unavailable.
• Final focused selection: 63 passed; original production and production-only revert each reproduce the same 13 failures with 50 passing controls.
• humanize.time statement coverage: 100%.
• Unchanged configured prek hooks, Ruff, Black, strict documentation build, and serial mypy pass. Default multi-worker mypy is blocked by sandbox IPC sockets.

No Windows/macOS/other-Python matrix or hosted-CI result is claimed. Benchmark functions execute with timing disabled, so this is not a performance measurement.

AI assistance: implementation, tests, and this description were prepared with OpenAI assistance.