malkhut(fix): wire workers into CMA training loop

CMAESTrainer.train() now accepts workers parameter and passes it to
evaluate_candidate(), enabling parallel episode evaluation during
actual training (not just in tests/benchmarks).

Benchmark result: ProcessPoolExecutor is optimal (4.76x speedup).
Ray is slower (0.36x) due to head init + plasma overhead for 90 scenarios.

MALKHUT/malkhut/training/cma_trainer.py

@@ -1237,11 +1237,13 @@ class CMAESTrainer:
         evaluator: PolicyEvaluator,
         pool: SelfPlayPool,
         store: Optional[MalkhutCHStore] = None,
+        workers: int = 0,
     ) -> None:
         self.codec = codec
         self.evaluator = evaluator
         self.pool = pool
         self.store = store
+        self._workers = workers
 
     def train(
         self,
@@ -1283,7 +1285,7 @@ class CMAESTrainer:
                 candidate = self.codec.decode(x, version=f"cma_{time.time_ns()}_{evals}")
                 score, results = self.evaluator.evaluate_candidate(
                     params=candidate, scenarios=scenarios, rng_seed=seed + evals,
-                    planner_type=planner_type,
+                    planner_type=planner_type, workers=self._workers,
                 )
                 perf_vec = PolicyEvaluator.performance_vector(results)
                 snap = PolicySnapshot(