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

astrai/inference/core/__init__.py

@@ -7,6 +7,7 @@ from astrai.inference.core.cache import (
     PagePool,
     RadixCache,
     ReqToTokenPool,
+    TaskCacheManager,
     page_hash,
 )
 from astrai.inference.core.executor import Executor
@@ -21,6 +22,7 @@ __all__ = [
     "PagePool",
     "RadixCache",
     "ReqToTokenPool",
+    "TaskCacheManager",
     "page_hash",
     "Executor",
     "InferenceScheduler",