Fix memory leak in RL training loop causing OOM at ~220GB

[Mr-Neutr0n]

Summary

Fixes #438

During RL training, system memory usage grows continuously until ~220GB and crashes. This PR addresses multiple memory leak sources across the store, daemon, trainer, and span exporter:

• InMemoryLightningStore: Added cleanup_finished_rollouts() method that removes completed rollout data (rollouts, attempts, spans, sequence IDs) and their associated tracking metadata (_completion_events, _start_time_by_rollout, _span_bytes_by_rollout, _running_rollout_ids, _evicted_rollout_span_sets) from all in-memory data structures after training data has been extracted.

• AgentModeDaemon.clear_data_and_server(): Now invokes store cleanup after clearing internal rollout tracking dicts, preventing rollout data from accumulating in the in-memory store across training steps.

• AgentLightningTrainer._train_step(): Added explicit del gen_batch after training data extraction, and gc.collect() + torch.cuda.empty_cache() at the end of each training step to release batch tensors and GPU cache.

• LightningSpanExporter: Added MAX_BUFFER_SIZE (10000) cap on the span buffer to prevent unbounded growth when spans fail to flush due to missing headers or unavailable store.

Test plan

• Run multi-step RL training and monitor memory usage via htop or /proc/meminfo - memory should stabilize rather than growing linearly
• Verify training metrics remain unchanged (cleanup happens after data extraction)
• Test with both v0 and v1 daemon modes to ensure no regressions
• Verify span exporter buffer cap works by checking warning logs when buffer fills

[Mr-Neutr0n]

Friendly bump! Let me know if there's anything I should update or improve to help move this forward.