fix: make timeline components hash deterministic

[FelipeDefensor]

Problem

TimelineComponentManager.hash_components concatenated component hashes in _components order, and restore_state recreated components by iterating a set of hash strings. Set iteration order over strings depends on PYTHONHASHSEED, so after an undo the components landed in _components in a process-dependent order and the digest changed even though the components were identical.

That order only matters because ORDERING_ATTRS is not a total order across component kinds. Staff orders by index and Clef by time, so a staff at index 0 and a clef at time 0 both have an ordinal of (0,) and neither sorts before the other. Clef ignores staff_index entirely, so the clefs of a multi-staff score all tie as well. bisect.insort_left leaves tied components in insertion order.

Worth flagging for reviewers: plain sorted(self._components) does not fix this. Timsort is stable, so tied components keep their input order and the digest still varies. A tiebreaker is required.

Impact: any undoable() test on a timeline with several components is flaky, and components_hash feeds the "is the file modified" comparison in are_tilia_data_equal.

Fix

• hash_components iterates sorted(self._components, key=lambda c: (c.ordinal, c.hash)) — total and reproducible, and identical to the ordinal order whenever ordinals are distinct.
• restore_state builds its delete/create lists by iterating the hash-keyed dicts rather than the sets, so the resulting component order is reproducible too.

Effect on hashes already saved in .tla files

• Timelines whose components have distinct ordinals (every kind other than score) hash byte-identically — verified against marker, beat, hierarchy and pdf timelines.
• Score timelines can hash differently — verified on a 27-component multistaff MusicXML score.
• This is harmless: the stored components_hash is discarded on load, since App._do_open overwrites file.timelines with freshly computed state. Planting a pre-fix hash in a .tla and reopening it leaves is_file_modified at False.

The digest's string shape (h1|h2|...|hn|, empty string when there are no components) is unchanged, and is now pinned by a test so it cannot drift silently.

Tests

Test	Pre-fix	Post-fix
test_score_timeline.py::TestHashComponents (2 tests)	fails on all seeds 0–5	passes on all
test_score_timeline_ui.py::TestClear::test_undo_redo	fails on seeds 0, 1, 4	passes on all
test_timeline_component_manager.py::TestHashComponents (2 tests)	passes (shape guard)	passes

The two test_score_timeline.py tests are the real regression guards: they are deterministic, so a single-seed CI run cannot pass them by luck. test_undo_redo reproduces the user-visible symptom, which is inherently seed-dependent.

Full suite: 1834 passed. The two tests/test_app.py failures under pytest -n auto are pre-existing xdist flakiness — dev fails a different pair the same way, and all 90 pass serially.