feat: add bias-aware GPS IMU fusion EKF

[zjn20030811]

Reference issue

Closes #603.

What does this implement/fix?

Adds a self-contained 2D GPS/IMU fusion example based on a bias-aware extended Kalman filter. The state estimates position, velocity, yaw, two accelerometer-bias terms, and gyroscope bias. Body-frame accelerations are rotated into the world frame, and bias random walks are included in the process covariance. GPS updates use a Joseph-form covariance update and an optional Mahalanobis innovation gate so isolated position outliers do not destabilize the estimate.

The existing CTRV example fuses velocity/yaw-rate observations but does not model raw accelerometer bias. This example complements it without changing existing APIs.

Additional information

• Added a deterministic curved-trajectory simulation and a localization documentation page.
• Added regression coverage for straight-line propagation, periodic yaw boundaries, finite-difference Jacobians, outlier rejection, covariance symmetry/positive semidefiniteness, and deterministic GPS improvement over inertial dead reckoning.
• Validation: python -m pytest tests -l -Werror --durations=0 (143 passed) on Python 3.12 with the repository dependency versions; Ruff and Black checks pass.

CheckList

• Added an unittest for the new algorithm.
• Added documentation for the new example.
• All local CI tests are green.

[zjn20030811]

Validation update: the EKF regression suite covers biased GPS/IMU innovations, Joseph-form covariance updates, and gating of rejected measurements. The implementation keeps the error-state covariance positive semidefinite while preserving the existing API. Focused tests and static checks pass.

@AtsushiSakai please review when convenient.