refactor: harden inference cache state and attention dispatch

- split KVCache into phase-specific PrefillKVCache/DecodeKVCache types selected by start_pos
- unify steady-state detection in TaskCacheManager
- guard decode steady-state reuse with the cached task signature so recycled req slots cannot replay a prior generation's tokens and positions
- collapse attention backend fwd_decode/fwd_prefill into a single subclass-owned forward with a shared _check_fwd guard
- fix thread-safety gap in weight update and validate prefill inputs before KV allocation
- centralize magic constants in InferenceConfig and align docs with behavior

astrai/inference/task.py

@@ -18,6 +18,7 @@ from typing import (
 
 from tokenizers.decoders import DecodeStream
 
+from astrai.config.inference_config import InferenceConfig
 from astrai.inference.metrics import MetricsCollector
 from astrai.tokenize.tokenizer import AutoTokenizer
 
@@ -25,6 +26,7 @@ if TYPE_CHECKING:
     from astrai.extension import AttentionBackend
 
 STOP = object()
+_config = InferenceConfig()
 
 
 @dataclass(frozen=True)
@@ -85,7 +87,7 @@ class Task:
         top_p: float = 1.0,
         top_k: int = 50,
         frequency_penalty: float = 0.0,
-        rep_window: int = 64,
+        rep_window: int = _config.default_rep_window,
         backend: Optional["AttentionBackend"] = None,
     ):
         self.task_id = task_id