docs: sync model and detector references

[AyushmanRaha]

Motivation

• Keep user-facing documentation aligned with the current registry and runtime code for supported MiDaS/DPT models and the 3D-tab detector.
• Clarify ONNX optionality and local-only detector behavior so setup/build guidance and API docs match code.

Description

• Added a short ## Model and detector specifications section to README.md (placed after "What I built vs third-party components") listing midas_small, dpt_hybrid, and dpt_large with canonical IDs, ONNX filenames, and input sizes.
• Updated docs/models-colormaps-metrics.md to include ONNX filenames and to state that ONNX is optional for standard PyTorch usage.
• Clarified POST /detect and /api/detect docs in docs/api-reference.md to explicitly identify the local TorchVision COCO detector (fasterrcnn_mobilenet_v3_large_320_fpn), note the internal 960 px cap, response metadata (detections, model, device_used, latency_ms, resolution), MPS→CPU resolution, and CPU retry behavior.
• Tightened docs/security.md wording about first-time model/detector weight caching and confirmed raw image payloads are not logged by telemetry.
• All changes are documentation-only and do not modify runtime code, frontend, Electron packaging, or CI workflows.

Testing

• Ran scripts/ci.sh docs-contract and python scripts/validate_docs.py; both passed.
• Ran scripts/ci.sh workflow-policy which passed.
• Ran scripts/ci.sh all partially: workflow-policy, docs-contract, black, ruff, and mypy succeeded, while pytest collection failed in this environment due to missing runtime/test dependencies (fastapi, numpy, pydantic, torch), which are expected to be present in a full test environment; this is an environmental limitation and not a code issue.
• Confirmed git diff --check shows no spacing or trailing-whitespace problems and committed the changes as a documentation-only update.

---

Codex Task