[XPU] Fix NemotronH (hybrid mamba2) launch on --device xpu (#32227)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

python/sglang/srt/mem_cache/memory_pool.py

@@ -2303,6 +2303,13 @@ class MHATokenToKVPool(KVCache):
             self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
         return self._get_value_buffer(layer_id)
 
+    def get_v_head_dim(self):
+        # Every layer in this pool is full-attention, so the value head dim is
+        # uniform and known at construction. Mirrors HybridLinearKVPool's
+        # get_v_head_dim() so the TritonAttnBackend mambaish branch works when a
+        # mamba2 config is served by a plain MHA pool (no per-linear-layer split).
+        return self.v_head_dim
+
     def get_kv_buffer(self, layer_id: int):
         return self.get_key_buffer(layer_id), self.get_value_buffer(layer_id)