fix(calendar): resolve Microsoft series masters through their calendar

[Chadpiha]

What & why

Customer report (v1.9.0): recurring Microsoft Calendar meetings render as "Untitled Event"
in Upcoming Meetings while one-off meetings sync fine.

Graph's calendarView/delta returns recurring-series occurrences as bare
{ id, type, seriesMasterId, start, end } stubs. #1665 backfills them from their series
master via GET /me/events/{id} — but that path only resolves items in the user's own
mailbox. For events on a shared calendar it 404s on every sync, so every recurring meeting
there is stored with a NULL summary, permanently. Two compounding gaps kept even transient
failures visible: incremental delta never re-delivers an unchanged occurrence, and a failed
backfill was never retried, so a bare stub stuck as an untitled block until the 7-day delta
token expiry.

Three changes, each covered test-first:

1. Calendar-scoped master fetch — /me/calendars/{calendarId}/events/{id}. A series
   master always lives in the same calendar as its occurrences, so the scoped path works for
   own and shared calendars alike.
2. Backfill retry — when any occurrence is still stripped after backfill, the delta token
   TTL is capped at 10 minutes, so an early full sync retries the master fetch instead of
   leaving untitled blocks for up to a week.
3. One-time heal migration (user_version 2) — clears stored Microsoft sync tokens so the
   first launch on this build runs a full sync and re-fetches rows that earlier builds stored
   untitled. Without it, a still-fresh delta token would keep pre-fix untitled rows on screen
   for up to 7 more days after updating.

Testing

• Full suite: 2981 tests, 0 failures (177 skips are the known better-sqlite3 ABI skips; the
  9 calendarDatabase tests, including the new migration test, were additionally run locally
  under ELECTRON_RUN_AS_NODE against the Electron binary — 9/9 pass)
• All three behavior changes were watched fail red before implementation

Follow-ups / limitations

• Wants a 1.9.1 patch release — the customer-visible fix only lands with the migration
  (remember the NEXT_PUBLIC_RELEASE_VERSION bump on the website after publishing)
• _apiGet still has no Retry-After handling for Graph 429s; the shortened-TTL retry
  bounds the damage (untitled for minutes, not days) but a throttled initial sync isn't
  retried within the same sync pass
• Free/busy-only shared calendars are unaffected: Graph genuinely returns "Busy" as the
  subject there, which is correct behavior, not this bug