fix: detect media_metadata changes for save-changes prompt (#377)

[FelipeDefensor]

Closes #377.

Summary

`FileManager.is_file_modified` compared the current app state against `self.file.dict`, but for `media_metadata` both sides referenced the same live `MediaMetadata` object: `on_set_media_metadata_field` mutates `self.file.media_metadata` in place, and `get_app_state` takes `dict(self.file.media_metadata)` of that same mutated object. The comparison was always equal, so editing notes never set the file to "modified" and the save-changes dialog never appeared on close.

The reporter flagged "notes" specifically, but it's not unique to that field — the title, custom fields, etc. were broken in the same way. Verified manually before the fix:

```
Initial is_file_modified: False
After title change is_file_modified: False
After notes change is_file_modified: False
```

After the fix:

```
Initial is_file_modified: False
After title change is_file_modified: True
After notes change is_file_modified: True
```

Approach

Track a snapshot of `media_metadata` taken at the last save / file-load.

• `self._saved_metadata` is set by a property setter on `file`, so every place that reassigns the file (`init`, `_setup_file`, `save`, `update_file`, `app.on_open`'s `self.file_manager.file = file`) refreshes the snapshot for free — no need to thread an extra method-call through every call site.
• `is_file_modified` compares `dict(current_data["media_metadata"])` against this snapshot. The existing `are_tilia_data_equal` call is kept for the other fields (`timelines_hash`, `media_path`); only the metadata-comparison branch is broken.
• `on_player_duration_changed` and `set_media_duration` also update the snapshot, since those mutations are system-driven (the player reports the loaded media's length) and shouldn't show up as user edits — without this, the existing `test_is_file_modified_empty_file` assertion would flip to True after the test fixture sets the duration to 100.

Test plan

• New regression tests `test_is_file_modified_after_notes_edit` and `test_is_file_modified_after_title_edit` cover the case the reporter hit. Both fail without the fix and pass with it.
• All existing modification tests keep passing (`test_is_file_modified_modified_tile`, `..._removed_field_from_media_metadata`, `..._modified_media_path`, `..._not_modified_after_update`, `..._when_modified_timelines`, `..._empty_file`). Their flow already used different objects on each side of the comparison, so the snapshot path is harmlessly redundant for them.
• Targeted suites pass: `tests/file/` 20 passed, 1 skipped; `tests/test_app.py` 63 passed (1 unrelated sentry-logging test deselected).

Note for the reviewer

The pre-existing `media_path` comparison has the same structural shape — `on_player_url_changed` writes the live path to `self.file.media_path`, and `get(Get.MEDIA_PATH)` reads it from the player, so the two sides drift only because the player doesn't immediately propagate. It happens to work in current usage because of that timing, but it's fragile. Out of scope here — happy to follow up if you'd like the same snapshot treatment for it.

[azfoo]

Added some tests, don't know if their behaviour is expected. Invert their logic if they are!