Fix: init_hidden always returns None (Mamba manages state internally)

The fallback init_hidden was returning a zero tensor instead of None,
causing test_hidden_state_api to fail. The fallback's _selective_scan
already handles None by initializing zeros internally, so init_hidden
should always return None for consistent API behavior.

Author: sunghunkwag

core/ssm_mamba.py

@@ -338,14 +338,8 @@ def init_hidden(self, batch_size: int = 1) -> Optional[torch.Tensor]:
             batch_size: Number of sequences in batch
 
         Returns:
-            Zero tensor of shape (batch_size, d_inner, state_dim)
-            if using the fallback.
+            None (Mamba manages state internally via selective scan).
         """
-        if not self._using_official_mamba:
-            d_inner = int(self.expand * self.d_model)
-            return torch.zeros(
-                batch_size, d_inner, self.state_dim, device=self.device
-            )
         return None
 
     def forward(