refactor: 将 adapter 层的计费/模型抓取/行为变体能力下沉到 core.api_format 注册表

- 新增 core/api_format/capabilities.py，统一注册计费模板、模型抓取、
  total_input_context 计算和 provider behavior variant
- 新增 core/usage_tokens.py，抽取 cache token 解析逻辑到 core 层
- handler adapter 移除各自的 compute_total_input_context / fetch_models /
  BILLING_TEMPLATE 覆盖，改为委托 core 注册表解析
- provider/behavior.py 改为薄封装，底层委托 core registry
- 新增 tests/test_architecture_import_rules.py 架构导入约束测试
- 新增 tests/services/api_format/test_capabilities.py 能力注册表测试

Closes #207

Co-authored-by: AAEE86 <ppk0227@hotmail.com>

tests/services/test_provider_keys_key_command_service.py

@@ -178,11 +178,13 @@ def test_clear_oauth_invalid_response_invalidates_caches(
         sys.modules, "src.services.cache.provider_cache", fake_provider_cache_module
     )
 
-    fake_models_service_module = types.ModuleType("src.api.base.models_service")
+    fake_models_service_module = types.ModuleType("src.services.cache.model_list_cache")
     setattr(
         fake_models_service_module, "invalidate_models_list_cache", _fake_invalidate_models_cache
     )
-    monkeypatch.setitem(sys.modules, "src.api.base.models_service", fake_models_service_module)
+    monkeypatch.setitem(
+        sys.modules, "src.services.cache.model_list_cache", fake_models_service_module
+    )
 
     key = SimpleNamespace(
         oauth_invalid_at=datetime.now(timezone.utc),