perf: move Q-stats to epoch boundary only — zero sync in training loop

Per-step reduce_current_q_stats created GPU pipeline bubbles every 50 steps.
Each sync waited for the GPU to drain all pending graph replays before
the CPU could continue. Q-stats are now computed at epoch boundary only
(process_epoch_boundary already calls reduce_current_q_stats).

Training loop is now fully async: 178 graph launches queue in <1ms,
GPU executes back-to-back with zero pipeline stalls.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -1565,9 +1565,9 @@ impl DQNTrainer {
                 } // end guard frequency check
                 guard_total_us += guard_start.elapsed().as_micros() as u64;
 
-                // Q-stats + v_range adaptation every step.
-                // Cost: ~1 kernel (q_stats_reduce) + sync for 76B DtoH readback.
-                // q_out_buf is already populated by graph replay — no re-computation.
+                // Q-stats computed at epoch boundary only (process_epoch_boundary).
+                // Per-step Q-stats sync was creating GPU pipeline bubbles.
+                if false {
                 if let Some(ref mut fused) = self.fused_ctx {
                     if let Ok(stats) = fused.reduce_current_q_stats() {
                         self.cached_avg_q = stats.q_mean as f64;
@@ -1586,6 +1586,7 @@ impl DQNTrainer {
                         fused.update_eval_v_range(stats.q_mean, q_std, q_gap, per_branch_q_gaps);
                     }
                 }
+                } // end if false (Q-stats disabled per-step)
                 train_step_count += 1;
                 self.gradient_logging_step += 1;
             }