[Core] Add realtime AR-Diffusion tick sessions for LingBot World 2.0 (#5491)

Signed-off-by: zhangbin.327 <zhangbin.327@bytedance.com>
Signed-off-by: Jack47 <scsvip@gmail.com>
Co-authored-by: zhangbin.327 <zhangbin.327@bytedance.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

tests/diffusion/ar_diffusion/test_kv_orchestration.py

@@ -54,13 +54,15 @@ def test_full_request_lifecycle_and_eviction():
         used = {int(s) // kv.block_size for s in slots}
         assert kv.null_block_id not in used
         assert used <= set(block_table)
-        # Resident window stays bounded by window + the in-flight chunk.
+        # Allocation may transiently add one in-flight chunk.
         assert len(kv.window_block_ids(adapter)) <= kv.spec.window_chunks + 1
-        free_after.append(kv.manager.block_pool.get_num_free_blocks())
         kv.commit_chunk(adapter)
+        # A successful commit immediately prunes physical ownership.
+        assert len(kv.window_block_ids(adapter)) <= kv.spec.window_chunks
+        free_after.append(kv.manager.block_pool.get_num_free_blocks())
 
     # Pool memory plateaus once the window is full (eviction recycles blocks).
-    assert free_after[-1] == free_after[kv.spec.window_chunks]
+    assert free_after[-1] == free_after[kv.spec.window_chunks - 1]
 
     kv.end_request(adapter)
     assert kv.manager.block_pool.get_num_free_blocks() == free_total