Extend evaluator with render mode control

[daphne-cornelisse]

Problem

It can be useful to render based on a particular criterion. For example, instead of always displaying the first env, you may want to continually show the worst scenario in a batch of scenes. This was not possible because vec_log(), which is used to collect stats from the envs, aggregates across agents and envs, losing per-env granularity.

Solution

• Added env_log() C binding that returns statistics per environment without aggregating across envs.
• Added configurable render selection modes (first, random, worst_score, worst_collision) via render_select_mode in the eval config.
• For worst-case modes, the Evaluator runs a stats-only rollout first, selects the target env using env_log, then renders that env in a second rollout.

Pipeline

Before

Before, the eval pipeline looked like this:

After

Now, if the render mode is not fixed nor random, it looks like this:

Logging example