Fix for the multiple-click issue in SegNext (non-SAM AI masks)

Hello,

Issue: In darktable 5.6, the SegNext model (mask-object-segnext-b2hq) only works with a single click. Unlike SAM, it is not possible to add additional points to refine the mask using additional clicks.
Ro-models ot cause: The SAM decoder has a has_mask_input flag that tells it whether to use the previous mask as context for iterative refinement. The SegNext ONNX graph does not include this flag: the previous mask is always passed to the decoder, making it impossible to distinguish a new prediction from a refinement pass. On a second click, the decoder receives two conflicting signals (new point + obsolete previous mask) and is unable to interpret them.
Fix (10 lines in src/develop/masks/object.c): For SegNext, reset the previous mask to zero before each call to the decoder.

Conclusion: Only one change is needed: in “src/develop/masks/object.c,” two blocks in _run_decoder():
Block 1 — Conditional reset of prev_mask (beginning of _run_decoder)
Replace:

  if(gui->guipoints_count <= 1 && !d->has_selection)
    dt_seg_reset_prev_mask(d->seg);

with :

  if(dt_seg_supports_box(d->seg))
  {
    // SAM: use has_mask_input flag for iterative refinement
    if(gui->guipoints_count <= 1 && !d->has_selection)
      dt_seg_reset_prev_mask(d->seg);
  }
  else
  {
    // SegNext: no has_mask_input flag, always start fresh
    dt_seg_reset_prev_mask(d->seg);
  }

Block 2 — reset in the self-refinement loop (beginning of the for loop) Add just before float *new_mask = dt_seg_compute_mask(...):

if(!dt_seg_supports_box(d->seg))
  dt_seg_reset_prev_mask(d->seg);

What do you think?

Greetings from the Luberon,
Christian