[codex] fix orchestrator routed experts memory retention

[samsja]

Summary

Fixes bounded routed-experts memory retention in the orchestrator step loop.

The routed replay path copies each rollout's tokens["routed_experts"] payload into TrainingSample.routed_experts, but the original rollout sidecars stayed attached to train_rollouts until end-of-step cleanup. The results list from interleave_rollout(...) also kept packed samples alive after the orchestrator had already extracted train_examples.

This change:

• clears routed-expert sidecars from rollout trajectories after conversion into training samples
• deletes the intermediate results list once samples have been extracted
• includes filter_df and timing_df in the explicit per-step cleanup before malloc_trim(0)

This reduces per-step RSS retention and peak memory in router replay runs. It does not claim to fix every possible monotonic production leak; ZMQ backpressure and monitor futures remain separate things to inspect if RSS still slopes upward.

Validation

• uv run pytest tests/unit/orchestrator/test_trajectories.py tests/unit/orchestrator/test_batch.py -q
• uv run ruff check src/prime_rl/orchestrator/orchestrator.py src/prime_rl/orchestrator/trajectories.py
• uv run ruff format --check src/prime_rl/orchestrator/orchestrator.py src/prime_rl/orchestrator/trajectories.py
• synthetic RSS probe comparing pre-patch-like retention vs patched cleanup:
  • pre-patch-like retained about +53.3 MB after cleanup for the probe payload
  • patched cleanup returned to baseline (+0.0 MB mean over baseline)