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

tests/extension/test_backend.py

@@ -154,14 +154,18 @@ class _DummyBackend(AttentionBackend):
     def supports_call(self, q, kv_cache, attn_mask, is_causal, fwd) -> bool:
         return True
 
-    def fwd_decode(
-        self, q, k, v, kv_cache=None, layer_id=0, attn_mask=None, is_causal=False
-    ):
-        return q
-
-    def fwd_prefill(
-        self, q, k, v, kv_cache=None, layer_id=0, attn_mask=None, is_causal=False
+    def forward(
+        self,
+        q,
+        k,
+        v,
+        kv_cache=None,
+        layer_id=0,
+        attn_mask=None,
+        is_causal=False,
+        fwd=None,
     ):
+        self._check_fwd(fwd)
         return q