feat(backend): rfdetr-cpp native object detection + segmentation backend

[localai-bot]

Summary

Adds a Go native gRPC backend (backend/go/rfdetr-cpp/) that dlopens librfdetr.so (built from mudler/rf-detr.cpp at the pinned RFDETR_VERSION) via purego and exposes the rfdetr.cpp inference pipeline through LocalAI's existing Detect RPC.

Supports all 5 RF-DETR detection variants (Nano/Small/Base/Medium/Large) and 3 segmentation variants (SegNano/SegSmall/SegMedium) with F32/F16/Q8_0/Q4_K quantizations. Pre-built GGUFs ship at mudler/rfdetr-cpp-* on HuggingFace (32 models total).

Detection returns Bbox + class_name + confidence; segmentation also returns PNG-encoded per-detection masks via the rfdetr_capi_get_detection_{class_id,box,score,class_name,mask_png} accessor functions.

Wiring (full adding-backends.md checklist)

• backend/go/rfdetr-cpp/{main.go, gorfdetrcpp.go, CMakeLists.txt, Makefile, run.sh, package.sh, test.sh, .gitignore}
• Top-level Makefile: BACKEND_RFDETR_CPP, docker-build-rfdetr-cpp, .NOTPARALLEL, prepare-test-extra, test-extra
• .github/backend-matrix.yml: CPU + CUDA 12 + CUDA 13 + HIP + Vulkan + SYCL f16/f32 + L4T variants
• backend/index.yaml: &rfdetrcpp meta + 17 image entries (latest + development for each variant)
• .github/workflows/bump_deps.yaml: RFDETR_VERSION pin tracking against mudler/rf-detr.cpp main branch
• gallery/index.yaml: rfdetr-cpp-nano + rfdetr-cpp-base model entries
• core/gallery/importers/rfdetr.go: GGUF auto-routing for HF imports (routes `.gguf`-bearing repos to `rfdetr-cpp`, falls through to Python `rfdetr` backend otherwise)
• core/gallery/importers/rfdetr_test.go: 6-entry table-driven test + live HF cross-check against `mudler/rfdetr-cpp-nano`

The `scripts/changed-backends.js` registration was checked and is not needed — the existing `endsWith("golang")` branch already routes `./backend/Dockerfile.golang` entries to `backend/go/${item.backend}/`.

Test plan

• `cmake --build` produces `librfdetrcpp-{avx,avx2,avx512,fallback}.so` per CPU variant
• `nm -D librfdetrcpp-fallback.so` shows all 11 `rfdetr_capi_*` symbols exported
• `go build -o rfdetr-cpp ./backend/go/rfdetr-cpp/` produces a 16 MB gRPC server binary
• `go test ./core/gallery/importers/...` passes (283 specs, 0 failed, 28s)
• End-to-end through `POST /v1/detection`: `local-ai run` + `curl` returns 9 detections for `rfdetr-base-f16` on a COCO test image, with correct `class_name` + `confidence` + `bbox` fields
• End-to-end seg through same endpoint: 10 detections for `rfdetr-seg-nano-f16`, each with a valid PNG-encoded mask (signature `89 50 4E 47`, decoded silhouettes match the upstream PyTorch output at IoU > 0.99)

Upstream

C++ implementation at https://github.com/mudler/rf-detr.cpp (now public). The rt-detr.cpp upstream:

• Matches PyTorch on CPU on every variant tested (1.05–1.45× speedup on Nano…Medium at T=8 on Ryzen 9 9950X3D)
• Lossless F16 (max |Δscore| ≤ 0.006 vs F32 reference)
• Q8_0 keeps full accuracy at 3.1× compression
• Sub-pixel bbox match vs PyTorch (54/55 at IoU ≥ 0.95 across 7 COCO val2017 images)
• Mask IoU 0.99+ vs PyTorch for segmentation

Assisted-by: Claude:claude-opus-4-7 [Claude Code]