fix: synchronize stream before DtoH readback after CUDA Graph launch

CUDA Graph capture disables cudarc event tracking, so device_ptr()
cannot rely on write events for ordering. Without explicit sync,
the DtoH memcpy can get CUDA_ERROR_INVALID_VALUE because the
kernel writes haven't completed yet.

Add stream.synchronize() between graph launch and scalar readback.
This ensures kernel output buffers (total_loss, grad_norm) are valid
before the host reads them. The sync cost is negligible — it's
1 sync per training step (already bottlenecked by kernel execution).

Fixes H100 CI failure: all DQN pipeline tests returned
"DtoH grad_norm: CUDA_ERROR_INVALID_VALUE" consistently.

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

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -864,8 +864,15 @@ impl GpuDqnTrainer {
             self.submit_training_ops(online_dueling, online_branching)?;
         }
 
+        // ── Synchronize stream before readback ──────────────────────
+        // CUDA Graph replay and direct kernel launches are async. The stream
+        // must be synchronized before DtoH to ensure kernel writes are visible.
+        // Also required because graph capture disables event tracking, so
+        // cudarc's device_ptr() can't rely on write events for ordering.
+        self.stream.synchronize()
+            .map_err(|e| MLError::ModelError(format!("stream sync before readback: {e}")))?;
+
         // ── Gather 2 scalars to host: total_loss + grad_norm ────────
-        // Read each 1-element GPU buffer directly to host (2 × 4B DtoH).
         let mut loss_host = [0.0_f32; 1];
         self.stream.memcpy_dtoh(&self.total_loss_buf, &mut loss_host)
             .map_err(|e| MLError::ModelError(format!("DtoH total_loss: {e}")))?;