fix(service): route engine creation through create_engine factory

[mountarreat]

What does this change?

backend/service.py::_get_engine constructed CorridorKeyEngine
directly, bypassing the CorridorKeyModule.backend.create_engine
factory that the CLI path already uses. That left the repo with
two parallel construction paths for the same engine.

This PR routes the service through the factory. To keep the
change a pure architectural refactor with no user-observable
behavior change, model precision is made explicit at both call
sites:

• create_engine in CorridorKeyModule/backend.py gains a
  model_precision: torch.dtype = torch.float16 parameter,
  forwarded to CorridorKeyEngine in the Torch branch. The
  default matches the value previously hardcoded inside the
  factory, so the CLI call site in clip_manager.py:633 is
  unchanged.
• _get_engine in backend/service.py now calls
  create_engine(backend="torch", device=self._device, img_size=2048, model_precision=torch.float32). Explicit
  backend="torch" preserves current service behavior (Torch on
  every platform, no MLX auto-pick). Explicit
  model_precision=torch.float32 preserves the service's
  existing precision (FP32 weights plus mixed_precision=True
  autocast, reached implicitly before by omitting the kwarg and
  falling through to CorridorKeyEngine's own default at
  inference_engine.py:60).

The CLI and service therefore continue to run at different
precisions (FP16 vs FP32) as they have been. This is now
documented intent at the call sites instead of an implicit
divergence from the factory bypass. Whether to unify the two
paths on a single precision is a separate design call.

The old "Loading checkpoint: " info log in _get_engine
is dropped because create_engine already logs the checkpoint
name plus device.

How was it tested?

TDD: four new regression tests were added in
tests/test_backend.py and observed failing on main before
the fix.

• TestCreateEngineModelPrecision (3 tests): the factory
  exposes model_precision, defaults to torch.float16, and
  forwards the requested value to CorridorKeyEngine (patched
  and inspected).
• TestServiceEngineRouting (1 test): a bare
  CorridorKeyService() calls _get_engine() and the call
  routes through create_engine with backend="torch",
  device="cpu", img_size=2048, and
  model_precision=torch.float32.

One informative pre-fix failure: the service-routing test saw
_get_engine reach _discover_checkpoint and attempt a real
HuggingFace download, confirming the factory was being bypassed.

Verified locally on Windows with Python 3.13.11:

• uv run ruff format --check: clean
• uv run ruff check: clean
• uv run pytest -q --tb=short -m "not gpu": 367 passed, 1
  skipped (MLX), 4 deselected (GPU). The +4 passes vs upstream's
  363 are the 4 new regression tests.
• uv run corridorkey --help: exits 0
• The two warnings in the suite output (PytestConfigWarning: Unknown config option: env and a torch autocast warning on a
  no-CUDA machine) are both pre-existing on main and not
  introduced by this change.

Checklist

• uv run pytest passes
• uv run ruff check passes
• uv run ruff format --check passes