[mkantautas]

Fixes #584.

Time.diff() built both totals from hours, minutes and seconds only, so both operands' microseconds were dropped. For example, Time(12, 0, 0, 500000).diff(Time(12, 0, 1)) returned 1 second instead of 0.5. Time - Time uses diff() and had the same error. This adds the microseconds to both totals.

closest() and farthest() compared diff(...).in_seconds(), which truncates to whole seconds. Before this change they could pick the wrong candidate when sub-second parts mattered. For example, Time(12, 0, 0, 900000).closest(Time(12, 0, 1), Time(12, 0, 0)) returned 12:00:00, which is 0.9 seconds away, instead of 12:00:01, which is 0.1 seconds away. With precise differences, truncation would also turn distinct sub-second distances into ties. They now compare total_seconds(), in line with DateTime.closest() and DateTime.farthest(), which already compare full-precision distances. Equal distances still return the second candidate.

Tests:

• tests/time/test_diff.py and tests/time/test_sub.py cover differences within and across seconds, signed results, the minimum and maximum times, and standard-library time operands.
• tests/time/test_comparison.py covers closest() and farthest() with sub-second distances, both candidate orders and standard-library time candidates.
• Six existing diff_for_humans() tests now freeze the clock, matching the equivalent DateTime tests. With microsecond precision, a future comparison against a ticking clock can fall just short of a minute or an hour and show the preceding unit, as DateTime already does.

Validation:

• Without the fix, the six new difference and subtraction tests fail (65 passed, 6 failed).
• All results below come from one run with Python 3.13 on Linux and python-dateutil 3.0.0.
• On master (e06dace), two of the six new comparison tests fail: the cross-second closest() and farthest() cases (24 passed, 2 failed).
• With the fix, tests/time/test_diff.py, tests/time/test_sub.py and tests/time/test_comparison.py give 97 passed with the pure-Python backend. With the compiled extension, they also give 97 passed; that run used the extension module from the pendulum 3.3.0 wheel, since this change touches no Rust code.
• Ruff 0.14.11 check and format: clean.
• Full suite with the pure-Python backend: 1,882 passed, 3 skipped, 1 failed. The failure is tests/test_main.py::test_instance_with_aware_datetime_any_tzinfo, which expects +02:00 and gets Europe/Paris. It fails the same way on unchanged master with the same dependencies.

Pull Request Check List

• Added tests for changed code.
• Updated documentation for changed code. Not updated: this restores microsecond precision and adds no API.