refactor: decouple task cache from PagePool and unify steady-state detection

- TaskCacheRegistry -> TaskCacheManager (independent, held by scheduler)
- TaskCacheState co-locates 5 parallel dicts into one dataclass
- AllocationStrategy base class + PagedStrategy subclass (page_size is a parameter)
- _rollback() helper for unified cleanup (no duplicate free paths)
- Task._kv_len + prefill_done property (explicit, no output_tokens proxy)
- Steady-state detection single-sourced in TaskCacheManager.bind()
- PagePool is now pure physical layer (no task knowledge)
- Removed dead _page_to_hash dict in RadixCache

tests/inference/test_task.py

@@ -20,11 +20,12 @@ def test_task_default_status_is_pending():
 def test_task_next_pos():
     task = Task("id1", [1, 2, 3])
     task.input_tokens = 5
+    task.mark_prefill_done()
     assert task.next_pos == 5
-    task.output_ids.append(4)
-    assert task.next_pos == 5
-    task.output_ids.append(5)
+    task.advance_kv()
     assert task.next_pos == 6
+    task.advance_kv()
+    assert task.next_pos == 7
 
 
 def test_task_is_finished_max_tokens():