[jacoblambert]

Summary

Planning review of the backfilled 1.0 samples found that ego_pose.acceleration held the raw IMU sample (gravity included, with a bias of about +0.5 m/s² at rest), while their planner uses Autoware's gravity-free estimate (/localization/acceleration, RMSE 0.12 m/s² against d(speed)/dt).

Both are useful for planning team, so we add support for both in this PR.

• ego_pose.acceleration: (ax, ay, az) in the vehicle frame, without gravity: the localization estimate, what Autoware feeds the planner. This reverses the "gravity included" line from feat(schema): caller-given metadata in V0Migrator.migrate (backfill) #319.
• New ego_pose.imu: option[[float;6]] = (ax, ay, az, wx, wy, wz), the raw IMU in the vehicle frame: specific force with gravity (az ≈ +9.8 m/s² at rest) and angular velocity.
• Spec sentence: an ego_pose sits at a sample time; twist, acceleration, imu, geocoordinate and covariance hold the latest message at or before timestamp, best effort; the pose itself is interpolated.

0.x ego_pose.acceleration is dropped when loading or migrating a 0.x dataset. Checked on local datasets: where 0.x has it (22a86de5, 77f0c093), it is the raw Tamagawa IMU in FRD (az ≈ −9.8), so it is neither 1.0 field, and the IMU setup is not recorded in the dataset. V0Migrator writes a note; the backfill restores both fields from the bag.

Master sheet: ego_pose.imu added and acceleration reworded as proposed.

Tests

• imu with 3 values is rejected (test_invalid_record).
• A 0.x dataset with acceleration migrates with both fields null and the note.
• The sample dataset gets imu: null; its manifest hash is updated.

[ktro2828]

LGTM