Fix full-day Durations collapsing to zero on PyPy (#876)

[uttam12331]

Summary

Closes #876.

On PyPy, duration(days=1) (and any whole-day / full-value duration) returned an empty Duration() with total_seconds() == 0.

Root cause

On PyPy, Duration defines its own total_seconds() (the if PYPY: branch) that computes the total from self._years, self._months, self._weeks/self._remaining_days/self._days, self._seconds and self._microseconds.

But Duration.__new__ used self.total_seconds() to derive self._total:

total = self.total_seconds() - (years * 365 + months * 30) * SECONDS_PER_DAY

at a point before those _days/_seconds/… attributes are assigned (a few lines later). During construction they are still at their class defaults of 0, so on PyPy total_seconds() returns 0, _total becomes 0, and the subsequent normalization zeroes out the whole duration. On CPython the method isn't overridden, so timedelta.total_seconds() runs and the value is correct — which is why the bug is PyPy-only.

Fix

Derive the total from timedelta.total_seconds(self), which reads the already-initialized underlying timedelta value and is independent of the PyPy override. On CPython this is equivalent to the previous call, so behavior there is unchanged.

Tests

Added test_new_does_not_depend_on_total_seconds_override, which runs on CPython: it monkeypatches Duration.total_seconds to return 0.0 (mimicking the PyPy override reading not-yet-set attributes) and asserts duration(days=1)._total == 86400.0. It fails on master (_total == 0.0) and passes with this change. Full tests/duration/ suite passes (48 passed); ruff check/ruff format clean.