Precompute pinhole target coordinates in image warps - #4930

Supersedes #4930.

## Problem
`WarpImageBetweenCameras` and `WarpImageWithHomographyBetweenCameras`
call `Camera::CamFromImg` and `Camera::ImgFromCam` on every pixel. Both
go through the 19-way runtime `switch (model_id)` in
`sensor/models/runtime.h` (~24M dispatches for a 12 MP image), which
prevents the compiler from inlining the concrete camera model across a
scanline, hoisting row-invariant terms, constant-folding unit depth (`w
= 1`), and vectorizing the inner loop.

## Fix
Hoist the `CAMERA_MODEL_SWITCH_CASES` dispatch out of the inner `x` loop
into two scanline helpers (`CamFromImgRow` and `ImgFromCamRow`) in
`src/colmap/image/warp.cc`, and unify both warp functions into a single
`WarpBetweenCamerasImpl` (+30 net lines in `warp.cc`):
- Optimizes both `WarpImageBetweenCameras` and
`WarpImageWithHomographyBetweenCameras`, while deduplicating their
setup, scanline loop, interpolation dispatch, and rescaling.
- Accelerates both `CamFromImg` (target) and `ImgFromCam` (source) for
**all** camera models without leaking camera math or parameter indices
outside `sensor/models/`.
- Kept local to `warp.cc` rather than `sensor/models/runtime.h`: other
`CamFromImg` / `ImgFromCam` loops in the codebase
(`P3PEstimator::Residuals`, `HomographyMatrixRayEstimator::Residuals`,
`UndistortCamera`, `UndistortReconstruction`) operate on 100–10k points
(<0.5–3 ms) or 3D rays with `w != 1`, whereas scanline generation (`x +
0.5` with scalar `y`, unit depth `w = 1`, and separate input/output row
buffers) is significantly faster than generic in-place point-array
batching.

### Comparison with #4930
#4930 special-cases `PinholeCameraModel` in `WarpImageBetweenCameras` by
indexing `params[0..3]` directly. Compared to #4930, this PR:
- Keeps all camera parameter layout and projection math inside
`sensor/models/`.
- Works for all 19 camera models on both source and target cameras (not
only `PINHOLE` targets).
- Covers `WarpImageWithHomographyBetweenCameras` (stereo rectification)
as well.
- Delivers larger speedups even on `PINHOLE` targets by also eliminating
per-pixel dispatch and `w = 1` division in `ImgFromCam`.

## Benchmark
4032×3024 RGB image (12.2 MP), Release build (`g++ 15.2 -O3 -DNDEBUG`)
on Intel Xeon @ 2.20 GHz, best of 5 runs after warmup; all outputs are
bit-identical to `main`:

| Function | Cameras (`source -> target`) | Interpolation | `main` |
#4930 | **This PR** |
|---|---|---|---|---|---|
| `WarpImageBetweenCameras` | `OPENCV -> PINHOLE` | Bilinear | 1175.9 ms
| 908.7 ms (−22.7%) | **723.9 ms (−38.4%)** |
| `WarpImageBetweenCameras` | `OPENCV -> PINHOLE` | Nearest | 726.3 ms |
447.4 ms (−38.4%) | **274.4 ms (−62.2%)** |
| `WarpImageBetweenCameras` | `SIMPLE_RADIAL -> PINHOLE` | Bilinear |
1109.1 ms | 844.7 ms (−23.8%) | **660.3 ms (−40.5%)** |
| `WarpImageBetweenCameras` | `OPENCV -> SIMPLE_PINHOLE` | Bilinear |
1217.4 ms | 1180.9 ms (−3.0%) | **669.5 ms (−45.0%)** |
| `WarpImageWithHomographyBetweenCameras` | `OPENCV -> PINHOLE` |
Bilinear | 1159.3 ms | 1159.3 ms (0.0%) | **767.6 ms (−33.8%)** |

## Tests
- `Warp.CameraModelsMatchPerPixelProjection`
(`src/colmap/image/warp_test.cc`) checks bit-exact output against
per-pixel `Camera::CamFromImg` and `Camera::ImgFromCam` across `OPENCV
-> PINHOLE`, `SIMPLE_DIVISION -> FOV` (partial `ImgFromCam` failure),
and `SIMPLE_RADIAL -> EQUIRECTANGULAR` (partial `CamFromImg` failure)
for both interpolation modes, both direct and intermediate-resolution
warp paths, and both `WarpImageBetweenCameras` and
`WarpImageWithHomographyBetweenCameras`.
- `image/warp_test`, `image/undistortion_test`, `image/line_test`, and
`controllers/undistorters_test` pass.