[NPU] Adapt hicache for K3 hybrid models (#39415)

test/registered/unit/mem_cache/test_decode_radix_lock_ref.py

@@ -442,6 +442,10 @@ class TestDecodeLockRefScenarios(CustomTestCase):
         queue.tree_cache.dec_lock_ref = MagicMock()
         queue.req_to_token_pool = MagicMock()
         queue.req_to_token_pool.available_size.return_value = 1
+        # Non-hybrid pools have no mamba allocator; MagicMock would otherwise
+        # auto-create one and break the `available_size() <= 0` comparison in
+        # pop_preallocated.
+        queue.req_to_token_pool.mamba_allocator = None
         queue.req_to_metadata_buffer_idx_allocator = MagicMock()
         queue.req_to_metadata_buffer_idx_allocator.available_size.return_value = 1
         queue.token_to_kv_pool = MagicMock()

test/registered/unit/mem_cache/test_npu_mamba_async_layout.py

@@ -156,6 +156,7 @@ class TestNPUMambaAsyncConfig(unittest.TestCase):
             MambaPoolHost.__dict__["_copy_tensor_all_layers_lf_pf"], staticmethod
         )
 
+    @patch.object(mamba_pool_host, "transfer_mamba_state", None)
     def test_conv_only_load_skips_empty_temporal_component(self):
         pool = MambaPoolHost.__new__(MambaPoolHost)
         pool.layout = "page_first_direct"