fix: preserve microsecond precision in Duration for large values

[SAY-5]

Pull Request Check List

• Added tests for changed code.
• Updated documentation for changed code.

Fixes #332.

Duration derived its normalized microseconds/seconds/days fields from total_seconds(), which returns a float and loses precision for large values:

>>> pendulum.duration(microseconds=8999999999999999).microseconds
999998  # datetime.timedelta returns 999999

The normalization now works on exact integer microseconds taken from the parent timedelta fields, matching datetime.timedelta behavior. The same float-precision issue in AbsoluteDuration (used by Time.diff()) is fixed the same way.

No documentation change needed; this restores the documented microseconds semantics.

[codspeed-hq]

Merging this PR will not alter performance

✅ 1 untouched benchmark

---

_{Comparing SAY-5:fix-duration-microsecond-precision (fa6a987) with master (b99bd14)}