feat: add optional CUDA kernel system (csrc/) + fused GQA decode attention

Structure:
  csrc/               -- .cu sources + build.py registry
  astrai/extension/   -- compiled .so + __init__.py (import dispatcher)
  setup.py            -- CUDAExtension from csrc/build.py REGISTRY

Control: CSRC_KERNELS=true|false env var at install time.
Fallback: astrai.extension.available dict for runtime detection.

astrai/__init__.py

@@ -16,6 +16,7 @@ from astrai.dataset import (
     Store,
     StoreFactory,
 )
+from astrai.extension import available
 from astrai.factory import BaseFactory
 from astrai.inference import (
     GenerationRequest,

astrai/extension/__init__.py

@@ -0,0 +1,13 @@
+import importlib
+import logging
+
+logger = logging.getLogger(__name__)
+
+available: dict[str, bool] = {}
+
+for _name in ["gqa_decode_attn"]:
+    try:
+        importlib.import_module(f".{_name}", package=__package__)
+        available[_name] = True
+    except ImportError:
+        available[_name] = False