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.

[Wkhanex]

The new month/year branches include seconds but still discard delta.microseconds before formatting. At higher requested precision, a fractional remainder is therefore lost:

delta = timedelta(days=30.5, microseconds=999999)
precisedelta(delta, minimum_unit="months", format="%.12f")
# actual: "1 month"
# numeric duration / (30.5 * 86400): 1.000000379477

The analogous 365-day year case renders “1 year” instead of retaining the numeric 1.000000031710 remainder. Half-day extra-seconds controls correctly render 1.016393 months and 1.001370 years at six decimals. I used Python 3.12.14 against commit 9e3f2fc.

Could both new numerators include usecs / 1e6 before setting days/secs/usecs to zero, with high-precision regressions? This is a remaining sub-day precision case, not a claim about default two-decimal output. I used English-locale standalone checks against fetched PR source with only build-generated version metadata stubbed, not the full pytest suite.

[PDGGK]

Thanks for checking. The other coarse units also drop microseconds when they are the minimum unit: on main, precisedelta(timedelta(minutes=1, microseconds=999999), minimum_unit="minutes", format="%.12f") gives "1 minute", and likewise for hours and days. This PR brings months and years in line with those units by including the sub-day seconds; carrying microseconds through every coarse unit would be a separate change.