easy - [verbosity] - Reduce log verbosity

[felipemello1]

(1) sets debug=False in the reward. Users can look at the samples directly in wandb and printed on the console
(2) Fix warning from using old torch api

/home/felipemello/.conda/envs/forge/lib/python3.12/site-packages/torch/cuda/memory.py:491: FutureWarning: torch.cuda.reset_max_memory_allocated now 
calls torch.cuda.reset_peak_memory_stats, which resets /all/ peak memory stats.
  warnings.warn( in /home/felipemello/forge/src/forge/observability/perf_tracker.py