fix: drain CUDA errors between hyperopt trials to prevent cascade

If a trial corrupts GPU state (CUDA_ERROR_ILLEGAL_ADDRESS from extreme
training), the inter-trial sync now drains the error instead of
propagating it — preventing all subsequent trials from failing.

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

crates/ml/src/hyperopt/adapters/dqn.rs

@@ -2656,9 +2656,14 @@ impl HyperparameterOptimizable for DQNTrainer {
         // Synchronize the trial device's stream to ensure all async frees complete.
         // With forked streams sharing the same context, free_async returns memory
         // to the context's pool — available for the next trial's allocations.
+        // If sync fails (CUDA_ERROR_ILLEGAL_ADDRESS from corrupted trial), drain
+        // the error and continue — the next trial creates a fresh stream.
         if let MlDevice::Cuda { stream, .. } = &trial_device {
-            stream.synchronize()
-                .map_err(|e| MLError::TrainingError(format!("CUDA sync between trials: {e}")))?;
+            if let Err(e) = stream.synchronize() {
+                tracing::error!("CUDA sync between trials failed (draining): {e}");
+                // Drain context-level error to prevent cascade to next trial
+                let _ = stream.context().check_err();
+            }
         }
         info!("Resource cleanup complete");