fix: acquire write lock once per epoch, not per training step

The async write lock (self.agent.write().await) was acquired inside the
per-step training loop, causing 696s/epoch overhead from async yield on
every iteration — 178 steps × ~4s per lock acquisition. Actual GPU compute
was only 3.6s (graph replay at 2ms/step).

Fix: acquire the lock once before the loop. The training loop is the only
writer during this phase — no contention, no need to release between steps.

Expected: epoch time drops from 700s to ~5s.

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

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

@@ -1473,10 +1473,10 @@ impl DQNTrainer {
             }
         }
 
+        // Acquire write lock ONCE for the entire training loop — not per-step.
+        // Per-step lock acquisition was burning 696s/epoch due to async yield overhead.
+        let mut agent = self.agent.write().await;
         for _step in 0..num_training_steps {
-            // Single write lock for both sample + fused step — eliminates
-            // read→write lock cycling overhead (~43ms per step from contention).
-            let mut agent = self.agent.write().await;
 
             let sample_start = std::time::Instant::now();
             {