states.kalman.information

The information form carries the *inverse* covariance instead of the covariance: an information matrix ``Y = P^{-1}`` and information vector ``y = P^{-1} mu``. Its appeal is the update step, which becomes a plain *addition* — fusing a measurement only adds ``H^T R^{-1} H`` to ``Y`` and ``H^T R^{-1} z`` to ``y`` — so multiple independent cues combine without a matrix inverse per fusion. The predict step pays for that by being the awkward one (it needs a Woodbury identity to add process noise). The filtered mean ``mu = Y^{-1} y`` is kept in the primary field so the existing cost zoo can match on it. This implementation uses a random-walk model (``F = H = I``), which is the appropriate "no motion" assumption for appearance/kernel embeddings. Its posterior is the exact Gaussian posterior, identical to :class:`~unitrack.states.KalmanLinear` + :class:`~unitrack.states.KalmanUpdate`.