refactor(model): split backend.py into per-backend modules

[stanlrt]

Summary

Splits src/raitap/models/backend.py into one module per backend, matching the existing tree_backend.py / xgboost_backend.py pattern. Pure structural refactor, zero behaviour change.

• base_backend.py — ModelBackend ABC + shared _adapt_input_shape
• torch_backend.py — TorchBackend + _is_torchvision_detection_model, _move_tensors_to_device, _torch_hardware_label
• onnx_backend.py — OnnxBackend + _resolve_onnx_expected_shape, _NUMPY_DTYPES_BY_ONNX_TYPE, _onnx_hardware_label
• backend.py removed; all ~37 importers repointed (clean break, no shims); dead logger dropped
• model.py imports the new backend modules at runtime so @register still fires for .onnx/.pth/.pt/.ubj
• Docs repointed: architecture.md, adding-a-backend.md, models/README.md

Closes #335.

Checklist

• CI — Required checks are green
• Breaking changes — Not a breaking change: only internal raitap.models.* import paths move; no public package API, config, or file format changes.
• Contributor guide — I’ve read Pull requests and commit messages before requesting review.

Optional

• Issue (optional) — Closes Split models/backend.py into per-backend modules #335.
• Docs (optional) — architecture.md, adding-a-backend.md, models/README.md repointed.
• Tests (optional) — No new tests; pure move, covered verbatim by the existing backend/registration suite.