feat(vitmatte): object-aware tiling and multi-object matte path

[johnwbrisbin]

What this changes

nodes/mask_matting/matters/vitmatte_backend.py

Replaces the full-image single-pass inference with an object-aware tiled approach.

Why: VitMatte's fixed 512x512 input means large images currently get
downscaled before inference and upscaled after, losing edge detail precisely
where it matters most. Tiling processes the image at native resolution, one
512x512 patch at a time, covering only the trimap uncertainty band around each
object rather than the whole frame.

Changes:

• matte(): rewritten to locate the object bounding box, select the minimum
  set of 512x512 tiles that covers it with at least 64px overlap, run
  inference on each tile, and blend results back with cosine-weighted
  accumulation. Empty masks short-circuit without any model call.
• matte_multi(): new method for N-object batches. Tiles shared between
  non-overlapping objects are processed once with a merged trimap; per-object
  alphas are recovered by trimap-presence masking, then max-merged into a
  composite alpha.
• _matte_multi_frame(): core per-frame accumulation loop used by both paths.
• Tile helpers: _global_tile_starts_1d, _optimal_tile_starts_1d,
  _best_tile_starts_1d, _object_tile_positions. Per-object tile count is
  the minimum of the object-optimal layout and the global grid covering the
  same region.
• _blend_weight(): cosine ramp over the overlap band to suppress seams.
• _verify_tile_logic(): offline geometry smoke-test (call manually or in CI).
• NEEDS_TRIMAP removed: the trimap is derived internally per object from the
  coarse mask and edge_radius; callers no longer need to supply one.

nodes/mask_matting/node.py

Adds a use_multi gate in the matting section. When a segmenter returns
object_masks and object_boxes in its output dict (indicating it produced
per-object segmentation, e.g. SAM 3.1 with multiplex tracking), and the
matter backend exposes matte_multi, and the batch size is 1, the node routes
through matte_multi() instead of the single-object matte() call. All three
conditions must hold; otherwise the existing single-object path is used
unchanged. No change to RETURN_TYPES or any node interface.

Test plan

• Single-object workflow with a standard segmenter: verify use_multi
  stays false and output is unchanged from before.
• Large image (>512px in either dimension): verify tiled output matches or
  exceeds quality of the previous downscale/upscale path at object edges.
• Multi-object segmenter output (SAM 3.1 multiplex): verify use_multi
  activates and each object receives an independent alpha.
• Empty mask input: verify the short-circuit path returns a zero alpha
  without crashing.