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/__init__.py

@@ -38,6 +38,7 @@ from astrai.inference.core import (
     RadixCache,
     ReqToTokenPool,
     Task,
+    TaskCacheManager,
     TaskManager,
     TaskStatus,
     page_hash,
@@ -67,6 +68,7 @@ __all__ = [
     "PagePool",
     "RadixCache",
     "ReqToTokenPool",
+    "TaskCacheManager",
     "page_hash",
     "sample",
     "BaseSamplingStrategy",