Improve goal rendering and enrich evaluation replays

[Victorbares]

What

• Capture world-space goals and configured goal radius in replay data.
• Display cumulative and per-step reward components and reward coefficients.
• Track road IDs and highlight the selected agent’s current lane.
• Assign evaluation maps deterministically, respecting explicit map indices.

Why

Goals previously depended on captured policy observations, causing them to disappear or render
incorrectly. Evaluation map selection could also mismatch replay scenarios. These changes make replays
reliable and provide better debugging context for goals, lanes, and rewards.

Notes

The replay payload now includes goals_f32, rewards_f32, coefs_f32, road IDs, and per-agent goal radii.