Proposal: Incremental, torch-only SAM-3 integration (inference-first)

This issue proposes an incremental integration of SAM-3 (Segment Anything Model v3) into Kornia, following existing contrib model patterns (e.g. SAM v1, RT-DETR).

The goal is to avoid a single large PR and instead iterate via small, reviewable PRs, starting with inference-only functionality, as discussed with maintainers.

proposed solution :

Constraints:

• Torch-only dependencies (PyTorch + existing Kornia deps)
• Inference-first (no training code, datasets, or CLI tools)
• Incremental PRs, each self-contained
• CI-safe from the first PR
• No breaking changes to existing SAM (v1)

Design direction: Kornia already integrates SAM v1 using a modular structure (image encoder, prompt encoder, mask decoder, model wrapper). SAM-3 would follow a similar design under a separate sam3 namespace, reusing existing patterns such as ModelBase, configs, and SegmentationResults where applicable.

Proposed components:

• Image encoder
• Prompt encoder (points / boxes / masks)
• Mask decoder
• Model wrapper + config
• Optional pretrained weight loading
• Minimal docs and examples

Incremental plan:

• Phase 1: Core architecture (image encoder only)
• Phase 2: Prompt encoder + mask decoder
• Phase 3: Model wrapper & end-to-end inference (random weights)
• Phase 4: Pretrained weights (optional)
• Phase 5: Docs & minimal examples

Each phase would be proposed as a separate PR.