Add AnyCalib single-image camera calibration - #4682

Add --calibration/--calibration_model_path to evaluate.py, run calibration
in the extraction step only, and record calibration settings in the report
metadata.

- Fix wrong attribute name in the benchmark tests, which referenced
  `args.calibration` instead of `args.camera_calibration` and therefore
  failed, including the pre-existing panorama test.
- Create the calibrator lazily when the controller runs, mirroring the
  feature extraction and matching controllers, so that the network is not
  held in device memory during preceding pipeline stages.
- Skip images whose dimensions do not match their database camera, which
  would otherwise store focal length and principal point at the wrong
  scale, since only the fitted parameters are written back.
- Skip learned calibration in the automatic reconstruction pipeline if
  explicit camera parameters were provided, which the image reader writes
  to the database and which must not be overwritten.
- Re-validate the aggregated calibration and fall back to the closest
  single-image calibration, because the coefficient-wise median of
  individually valid calibrations is not itself guaranteed to be
  projection-stable for high-order distortion models.
- Reject implausible per-image predictions via HasBogusParams, with the
  bounds exposed as backend-agnostic calibration options.
- Define backend-specific options via shared_ptr, as for feature
  extraction.
- Reuse Median from colmap/math and remove the dead
  ReverseScaleAndShiftParams.

- Share IsValidCalibration between the per-image and the aggregation path,
  which previously used two divergent copies, so that neither applied the
  full validation. NOTE: per-image predictions are now also checked for
  projection stability.
- Pass the target camera model to AggregateCameraCalibrations explicitly,
  instead of requiring the caller to set it beforehand, and validate on a
  candidate, so that the camera is only modified on success.
- Compute the closest calibration in a single pass over the deviations.
- Use Camera::CreateFromModelId in the tests and merge duplicated cases.

The calibrator is created lazily in the controller's worker thread, where
an escaping exception terminates the whole process instead of failing the
stage. Previously it was created in the constructor on the main thread,
where it propagated normally. A corrupt or missing model file therefore
aborted the automatic reconstruction pipeline with SIGABRT after feature
extraction had already run.

Ceres terminates without taking a step when the closed-form init is already at the optimum (noise-free pinhole data), reporting bit-identical initial/final costs and breaking the strict < assertion on some toolchains. This matches the product's own final <= initial success contract.

Harden validation (non-throwing Check, IsValidCalibration edge cases,
tighter round-trip tolerance), validate ONNX dtypes in AnyCalib/ALIKED/LoMa
via new element-type accessors, harden ray fitting (LDLT definiteness,
inverse-focal epsilon, FOV bounds, invalid-prior tolerance), make the
calibration controller non-fatal, verify export weights hash and record
provenance, and extend tests and docs.

Extend the linear solvers in ray_fitting.cc with a fisheye variant that
fits the equidistant (theta) projection instead of X/Z, and dispatch
radial-fisheye models to it for their leading k1[, k2] coefficients.
Undistorted fisheye models use the equidistant pinhole fit. The FOV gate
uses an incidence-angle check for fisheye rays.

HasRadialDistortionPrefix now returns optional<pair<int, bool>> with the
coefficient count and fisheye flag instead of pointer out-params.

Tests: wide-FOV (~170 deg) round-trip cases for all fisheye models, plus
an init-only test (zero refinement iterations) pinning the closed-form
solution.

InitializeCameraParams now uses naive initialization (focal length from
the image span, principal point at the data center, zero distortion)
instead of the AnyCalib-style closed-form linear solvers. Stress tests
over strong distortion (incl. near-pole division), noise, sparsity,
partial coverage, and outliers showed the Ceres refinement reaches the
same optimum from the naive start every time, so the linear init only
saved iterations at the cost of ~270 lines of solver machinery.

Removes the now-unused max_fov_deg option (incl. the
CameraCalibration.max_fov_deg CLI flag). Tests: drop the obsolete
init-only test; FULL_OPENCV round-trip skips distortion-value assertions
since naive-start refinement may land on a projection-equivalent point
in the weakly identifiable subspace (projection check still binds).

- Widen before multiplication in ray index computation
- Add default cases to rot90 switches and drop const blocking move
- Keep intentional integer crop-offset division out of float context
- NOLINT intentional out-of-range enum cast and getenv in tests

- Wire --image_list_path through the camera_calibrator CLI and pass the selected image names to the calibration controller, as documented.
- Default CameraCalibration.camera_model to empty to preserve each database camera's existing model; an explicit value still converts all cameras. Reject invalid models gracefully before inference, since the camera-model switches throw on them.
- Standardize the new calibration files on SPDX license headers.