📦 NEW: filter output of sam and tansform to expected output

[violoncelloCH]

this is based on #55 , so #55 needs to be merged first

Using the the SAM model, this applies more or less sophisticated filtering to get rid of masks which are not buttons. This is done in a few steps:

1. Construct a tree based on the bounding boxes of the masks.
2. In this tree search masks which could plausibly represent the remote (or a bit more than the remote) based on a few criteria:
   • a fixed point close to the lower clamp which always is on the remote
   • an assumed minimum size for any remote
   • an assumed maximum size for any remote (based on the space constraints we have anyway
   • the assumption that a remote has at least 3 buttons
3. Out of the masks which possibly represent remotes, the mask with the highest stability is selected.
4. All childs of that masks are put back in a flat list.
5. Masks which are too big compared to the remote are considered not to be buttons and are ignored.