[Feature] KPM/NFT Tracking Port to Rust

[kalwalt]

feat: Port KPM/NFT tracking layer (FreakMatcher) to Rust

Issue body

Summary

This issue tracks the complete porting of the KPM (Keypoint Matching) / NFT
(Natural Feature Tracking) pipeline from C++ to Rust. KPM is the subsystem
responsible for marker-less image tracking using FREAK binary descriptors and
a Difference-of-Gaussians feature detector.

The work is organized as a Strangler Fig migration: a thin FFI bridge
keeps the existing C++ running while Rust replaces it layer by layer. At every
step, a numerical regression test suite validates that the Rust output matches
the C++ output to within floating-point tolerance.

---

Background

The C++ source lives in
WebARKitLib under
lib/SRC/KPM/ and lib/SRC/KPM/FreakMatcher/. The Rust port lives in
this repository (crates/kpm/).

What has already been ported (in crates/core/):

Module	File	Status
ICP pose estimation	src/icp.rs	✅ Complete
Matrix math (ARMat)	src/math.rs	✅ Complete
Core types (ARParam, ARdouble)	src/types.rs	✅ Complete

What this issue covers — the new crates/kpm/ crate:

Layer	C++ source	Effort
KPM types & structs	kpmType.h, kpm.h	Low
KPM public API	kpmHandle.cpp, kpmResult.cpp	Low
Image resize helpers	kpmUtil.cpp (genBWImage*)	Low
Ref dataset I/O	kpmRefDataSet.cpp	Low
kpmMatching orchestration	kpmMatching.cpp	Medium
Math utilities	math/indexing.h, math_utils.h	Low
Linear algebra & solvers	math/linear_algebra.h, linear_solvers.h	Medium
Homography estimation	homography_estimation/robust_homography.h	High
Hough voting + feature matcher	matchers/hough_similarity_voting.h	Medium
Binary hierarchical clustering	matchers/binary_hierarchical_clustering.h	High
FREAK descriptor	matchers/freak.h, keyframe.h	High
DoG detector + Gaussian pyramid	detectors/DoG_scale_invariant_detector.*	High
VisualDatabase + facade	matchers/visual_database.*, facade/	High

Key dependency on Eigen: Eigen is used in exactly one place —
IncrementalHomographyFromLieWeights in robust_homography.h —
for a 3×3 matrix exponential. This will be replaced with either
nalgebra's matrix-exp feature or a hand-written Padé approximation.
The original authors left a // TODO: remove Eigen comment there.

---

Architecture

// note this was refactored: Now we have only core and wasm crate.
// kpm is inside core -> core/kpm
crates/kpm/
├── build.rs              # compiles C++ FreakMatcher as static lib (FFI phase)
├── src/
│   ├── lib.rs
│   ├── kpm_c_api.h       # thin C wrapper over C++ KPM (FFI phase)
│   ├── kpm_c_api.cpp     # implementation of C wrapper
│   ├── kpm_ffi.rs        # bindgen-generated bindings (FFI phase)
│   ├── types.rs          # KpmHandle, KpmRefDataSet, FreakFeature, etc.
│   ├── backend.rs        # FreakMatcherBackend trait
│   ├── cpp_backend.rs    # CppFreakMatcher: implements backend via FFI
│   ├── handle.rs         # KpmHandle struct (orchestration)
│   ├── ref_data_set.rs   # KpmRefDataSet: generate, save, load, merge
│   ├── matching.rs       # kpm_matching() + kpm_util_get_pose_binary()
│   └── freak/            # (Milestone 3+) pure Rust FreakMatcher
│       ├── mod.rs
│       ├── math.rs
│       ├── homography.rs
│       ├── hough.rs
│       ├── clustering.rs
│       ├── descriptor.rs
│       ├── detector.rs
│       └── database.rs
└── tests/
    └── regression.rs     # numerical regression vs C++ baseline

Cargo features

Feature	Purpose
ffi-backend (default)	Enables C++ FreakMatcher via FFI
dual-mode	Runs Rust and C++ in parallel and logs divergences

When ffi-backend is disabled and the pure Rust FreakMatcher is complete,
the C++ dependency is dropped entirely.

---

How to contribute

1. Pick a task from the milestone that interests you.
2. Each task has a corresponding prompt in CLAUDE_CODE_PLAN.md
   at the repo root — feel free to use it with Claude Code.
3. Every new function must have a unit test and must not break
   cargo test -p kpm.
4. Numerical functions must pass the regression tests in
   crates/kpm/tests/regression.rs.