ImportTask.chosen_info() was returning self.match.info.copy() for APPLY
actions, which preserves raw AlbumInfo/TrackInfo field names (album_id,
track_id) instead of the library model names (mb_albumid, mb_trackid).
find_duplicates() then built a temporary Album/Item with mb_albumid
unset, causing the duplicate query to match library entries with an
empty MBID rather than the real duplicate.

Fix chosen_info() to return dict(self.match.info.item_data) for APPLY
actions. item_data applies MEDIA_FIELD_MAP which translates all
AlbumInfo/TrackInfo field names to their library equivalents, so
find_duplicates() now builds the temporary model with the correct
fields including mb_albumid and mb_trackid.

Also update ImportTask.find_duplicates() to handle the renamed artist
field: AlbumInfo.item_data maps 'artist' -> 'albumartist', so the
albumartist override is now guarded by a check for the 'artist' key
(present for ASIS/RETAG paths, absent for APPLY after item_data).

Add regression tests for both album (mb_albumid) and singleton
(mb_trackid) duplicate key paths.

Fixes beetbox#7067

When AlbumInfo.artist is None, item_data omits 'albumartist' entirely
(None values are filtered during item_data construction).  The guard
  if info.get("albumartist") is None: return []
was previously uncovered.  Add a focused unit test that constructs an
APPLY task with an artist-less AlbumMatch and asserts find_duplicates
returns [] immediately.