fix(managers): ManagerBasedRVEnv.reset accepts seed (honor env.reset(seed=) contract) (#800)

ManagerBasedRVEnv (the manager-based RL base for mjlab v2 / Isaac-Lab-style
tasks) overrode reset(self, env_ids=None) without seed, so gym.make(task).reset(seed=)
raised 'TypeError: reset() got an unexpected keyword argument seed' for every
manager-based task — surfaced by the MuJoCo benchmark (12 mjlab v2 tasks failed
to reset). Add seed=None and forward to handler.set_seed (this base reimplements
reset and doesn't call super), matching the seed-contract fix already applied to
the other reimplementing bases (humanoid LeggedRobotTask, beyondmimic).

Verified on MuJoCo (8xRTX5090): mjlab.cartpole_swingup_v2 / cartpole_balance_v2 /
velocity_flat_g1_v2 now load + reset(seed=0) + run a rollout (previously TypeError).

Author: geng-haoran

roboverse_learn/managers/manager_based_rv_env.py

@@ -336,8 +336,18 @@ def _apply_setup_events(self) -> None:
     # env api
     # ------------------------------------------------------------------
 
-    def reset(self, env_ids: Sequence[int] | None = None):
-        """Reset selected envs (or all)."""
+    def reset(self, env_ids: Sequence[int] | None = None, seed: int | None = None):
+        """Reset selected envs (or all).
+
+        ``seed`` is forwarded to the handler when supported, honoring the
+        ``env.reset(seed=)`` contract (this base reimplements ``reset`` and does
+        not call ``super().reset``). Without this, ``gym.make(...).reset(seed=)``
+        raised ``TypeError`` for every manager-based (mjlab v2) task.
+        """
+        if seed is not None:
+            set_seed = getattr(self.handler, "set_seed", None)
+            if callable(set_seed):
+                set_seed(seed)
         if env_ids is None:
             env_ids = torch.arange(self.num_envs, dtype=torch.int64, device=self.device)
         elif not isinstance(env_ids, torch.Tensor):