fix: hyperopt preload buffer_size=1 → batch_size (PER capacity floor)

The data preload created a dummy DQNTrainer with buffer_size=1, which
fails the GPU PER capacity check (capacity must be >= batch_size).
This caused the preload to fail silently, falling back to per-trial
data loading from disk — 50 × 5s = 250s wasted.

Fix: set buffer_size = max(batch_size, 1024) so the PER allocation
succeeds. The preload trainer doesn't train — it just loads data into
a shared Arc for all trials.

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

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

@@ -878,7 +878,10 @@ impl DQNTrainer {
             let mut preload_hyperparams = DQNHyperparameters::default();
             preload_hyperparams.mbp10_data_dir = self.mbp10_data_dir.clone().unwrap_or_default();
             preload_hyperparams.trades_data_dir = self.trades_data_dir.clone().unwrap_or_default();
-            preload_hyperparams.buffer_size = 1;
+            // Buffer size must be >= batch_size for GPU PER allocation.
+            // Preload doesn't train — just loads data — but the trainer
+            // constructor still creates a PER buffer.
+            preload_hyperparams.buffer_size = preload_hyperparams.batch_size.max(1024);
             let mut loader = InternalDQNTrainer::new_with_device(preload_hyperparams, self.device.clone())
                 .map_err(|e| MLError::TrainingError(format!("Failed to create data loader: {e}")))?;