fix: return one calendar event per lesson, not one per teacher

[nickknissen]

Problem

Aula's calendar returns a separate lesson row for every adult attached to a
lesson. The rows carry identical title, startDateTime, endDateTime,
primaryResource and belongsToProfiles, differing only in id and the
teacher in participants.

get_calendar_events mapped each row straight to a CalendarEvent, so a
lesson with two adults — common in the younger years — came back twice:

09:30  DFA 2, Lærer Nummer 1
09:30  DFA 2, Lærer Nummer 2
10:15  DFA 2, Lærer Nummer 1
10:15  DFA 2, Lærer Nummer 2

Every consumer inherited the duplicates, all day, every day.

Change

merge_duplicate_lessons() folds those rows back into one event each,
collecting the adults into new teacher_names / substitute_names lists.

The merge key is deliberately conservative — title, timeslot, room and
child must all match. Two different subjects in the same slot, the same subject
in two periods, and a lesson shared across two children all stay separate, so
nothing real is hidden. The helper returns new events via dataclasses.replace
rather than rewriting its inputs, keeps the first row's id and _raw, and
sets has_substitute when any row reported one.

_teacher_names() reads every participant in a role instead of just the first.
Aula normally splits co-taught lessons across rows, but a single row carrying
several same-role participants has also been seen; handling both shapes means
the teacher list is complete either way.

Compatibility

teacher_name and substitute_name are unchanged and point at the first
adult, so existing callers keep working without edits. The new lists are
additive and appear in dict(event) for JSON consumers.

One follow-on was required rather than optional: both CLI summaries rebuilt
their teacher list from the duplicate rows (dict.fromkeys(ev.teacher_name ...)).
Left alone they would have regressed to naming a single adult once the rows
merged, so they now read teacher_names through a shared _unique_names
helper — which also removes a copy-pasted comprehension.

Tests

11 model tests covering the merge and each case that must not merge, plus two
client-level tests: the real duplicated-row payload, and multiple teachers on a
single row.

813 passed, 4 skipped; ruff check and ruff format clean.