feat: support H5 evaluation dataset export

[xtk8532704]

Summary

Adds the rosbag-to-H5 export path required to build closed-loop Diffusion Planner evaluation datasets.

Before frame extraction, each source bag is normalized to the current Autoware message schemas with autoware_msg_bag_converter and cached. The exporter then writes schema-v4 H5 shards and a Parquet frame index containing the pose and route metadata needed by closed-loop replay.

This PR replaces #395 and is based on new-architecture/main.

Changes

Rosbag schema normalization

• Runs every source bag through autoware_msg_bag_converter before ml_planner_data reads it.
• Caches normalized bags under the configurable converted_bag_root.
• Writes conversions through an .incomplete directory and publishes them atomically after validating that metadata.yaml exists.
• Adds and pins the converter, autoware_auto_msgs, and ros2_numpy workspace dependencies.
• Exposes the converter's vendored autoware_internal_planning_msgs_v1_13 and autoware_perception_msgs_v1_7 packages to the ROS workspace through top-level symlinks, allowing old CDR layouts to be deserialized during conversion.
• Keeps defensive per-message deserialization handling in ml_planner_data: remaining incompatible messages are reported as warnings, and dropout validation is disabled only when every message for an optional topic is unavailable.

Closed-loop dataset export

• Adds the override source split and its map-directory layout.
• Follows directory symlinks while discovering staged rosbag selections.
• Adds configurable future_steps (default: 80) and uses it consistently for label generation, valid-frame bounds, and topic-gap exclusion ranges.
• Extracts ego_x, ego_y, and ego_yaw from odometry and stores them in both H5 metadata and the Parquet index.
• Adds optional split_routes generation:
  • consecutive route messages with the same start pose form one route group;
  • minimum travel distance is evaluated per route group;
  • each retained group is written to its own H5 shard;
  • route_group_id is included in shard metadata and the Parquet index.
• Adds the Aaq3qv2Q / ePalette vehicle dimensions used by the evaluation data.

H5 output and documentation

• Introduces a shared schema-v4 H5 writer for full-bag and single-frame/open-loop exports.
• Validates the shared frame axis, finite numeric values, and strictly increasing timestamps.
• Writes H5 files atomically and chunks model tensors one frame at a time.
• Extends the ml_planner_data Python API with per-frame metadata access for shared open-loop output.
• Tightens resume validation so incompatible or incomplete shards are rejected rather than silently reused.
• Documents the H5/Parquet schema and the adjacent frames.tags.json sidecar format for mutable segment tags.
• Documents ROS dependency import/build steps and the required frame interval for each workflow.

Defaults and compatibility

• H5 format version is now 4. Existing version-3 shards must be regenerated with overwrite=true; resume=true rejects older shards.
• future_steps=80 preserves the existing prediction horizon by default.
• split_routes=false preserves one-shard-per-bag behavior for ordinary train/validation generation.
• Training and open-loop generation default to frame_interval=0.5.
• Closed-loop releases must set frame_interval=0.1 to match the 10 Hz rollout clock.
• Converted bags consume additional disk space under converted_bag_root and are reused on subsequent runs.

Validation

• Built ml_planner_data in the ROS 2 workspace with the pinned dependencies.
• Verified that colcon discovers the converter and both vendored legacy message packages.
• Verified Python syntax for create_h5_dataset.py.
• Checked modified C++ sources with clang-format --dry-run --Werror.