Add AGENTS.md file - #4162

## Summary

Adds a reusable [Agent Skill](https://agentskills.io)
(`.agents/skills/colmap-review/SKILL.md`) that instructs coding agents
(Codex, Gemini CLI, Cursor, Claude Code, etc.) to perform
maintainer-grade code reviews on COLMAP changes — uncommitted diffs,
commits/branches, or GitHub PRs — without modifying code.

This complements `AGENTS.md` (#4162): agents load only the short
frontmatter description by default and pull in the full review
instructions on demand (e.g. `"review this"` or `"review my PR"`), and
`AGENTS.md` links to it for harnesses without native skill discovery,
keeping regular coding sessions lean.

### What it covers
- **Modularity, readability & diff size:** Prefers short, cohesive
functions with self-documenting names and tightly scoped (`const`)
variables over long functions with wide scopes or opaque blocks, checks
whether sub-steps can be extracted into helpers or already exist in
COLMAP (`math/`, `geometry/`, `sensor/`, `scene/`, `estimators/`),
removes dead code and speculative options, and pushes scattered
`if/else` plumbing behind existing domain abstractions (`Camera`, `Rig`,
`Frame`, `TwoViewGeometry`).
- **Geometric, numerical & SfM/MVS correctness:**
Non-perspective/fisheye/spherical camera rays (`CamRayFromImg` vs.
`CamFromImg`), pixel-center (`+0.5`) vs. integer-grid coordinate
conventions, panoramic seam wrap-around, solver
conditioning/degeneracies (including on the final RANSAC inlier set) and
convergence/`allFinite()` checks, RANSAC vs. Ceres outlier/cheirality
residual semantics, missing Ceres manifolds (`EigenQuaternionManifold`,
`SphereManifold`), `target_from_source` frame and `Rig` composition, and
`FlatHashMap` iteration nondeterminism.
- **Safety, `pycolmap` & CI portability:** Precondition checks before
mutating `Reconstruction`/`Database`, clean exception propagation
(`THROW_CHECK*`) over `LOG(FATAL)` in library/worker code, schema/format
backwards compatibility, `pycolmap` container lifetime/GIL/STL
copy-conversion/`None`/option-parity hazards, and
feature-gate/CUDA/HIP/headless/Windows/formatting portability.
- **Performance & empirical evidence:** `O(N^2)` patterns, hot-loop
copies/allocations, and requiring benchmark justification (e.g. ETH3D or
multi-seed synthetic sweeps) for algorithmic or default-parameter
changes.
- **Tests & conventions:** Focused C++ regression tests (`*_test.cc`)
over bloated `pycolmap` pipeline tests, catching vacuous assertions,
reusing COLMAP matchers (`ReconstructionEq`, `Rigid3dNear`,
`SynthesizeDataset`, in-memory SQLite), and checking concrete repo
conventions (`FlatHashMap`/`NodeHashMap`, strong IDs,
`target_from_source`, SPDX headers).
- **Actionable output contract:** Every finding must be verified against
connected callers in the repo and have either a concrete failure trigger
or a concrete simplification, grouped into `Must fix` / `Should fix` /
`Nit` with a diff-reduction summary and verdict.