refactor: unify kernel module loading and packaging

- loader.py: lazy/cached import; is_available defers the actual load; get_module raises on unavailable
- ops/{attention,rotary,fp8}: use get_module instead of touching private _modules or their own _mod() cache
- package-data: ship astrai.extension.lib *.so in built wheels (non-editable installs previously lost every kernel)

astrai/extension/loader.py

@@ -5,9 +5,15 @@ Each kernel is built by the CMake build in ``csrc/CMakeLists.txt`` into a
 ``.so`` name equals the pybind name (e.g. ``attn_decode``, defined via
 ``TORCH_EXTENSION_NAME``). ``KERNEL_NAMES`` is discovered automatically from
 the ``.so`` files present, so adding a kernel to the CMake ``KERNELS``
-registry needs no change here. On import we try to load each one; kernels
-that failed to build (or are running on a CPU-only machine) are marked
-unavailable so the wrapper functions can fall back to ``torch`` SDPA.
+registry needs no change here.
+
+Loading is **lazy and centralized**: module names are discovered eagerly
+(cheap glob), but each ``.so`` is imported on first use via the single
+``get_module`` accessor, then cached. The wrapper modules (``ops/*.py``) never
+touch the internals or keep their own caches — they call ``get_module(name)``
+(or ``is_available(name)`` when a torch fallback is acceptable). A kernel that
+failed to build (or is running on a CPU-only machine) is ``None`` in the cache,
+so ``is_available`` returns ``False`` and ``get_module`` raises a clear error.
 """
 
 import glob
@@ -34,21 +40,44 @@ KERNEL_NAMES = _discover_kernel_names()
 _available: dict[str, bool] = {}
 _modules: dict[str, object] = {}
 
-for _name in KERNEL_NAMES:
-    try:
-        _mod = importlib.import_module(f".lib.{_name}", package=__package__)
-        _available[_name] = True
-        _modules[_name] = _mod
-    except ImportError:
-        _available[_name] = False
-        _modules[_name] = None
+
+def _try_load(name: str) -> object:
+    """Import and cache the ``name`` kernel module (lazy, one attempt).
+
+    Returns the module, or ``None`` if it is unavailable. Cached so each
+    ``.so`` is imported at most once per process.
+    """
+    if name not in _modules:
+        try:
+            _modules[name] = importlib.import_module(
+                f".lib.{name}", package=__package__
+            )
+            _available[name] = True
+        except ImportError:
+            logger.warning("kernel '%s' failed to import; marking unavailable", name)
+            _modules[name] = None
+            _available[name] = False
+    return _modules[name]
 
 
 def is_available(name: str) -> bool:
-    """Return ``True`` if the compiled kernel ``name`` was loaded."""
+    """Return ``True`` if the compiled kernel ``name`` could be loaded."""
+    if name not in _available:
+        _try_load(name)
     return _available.get(name, False)
 
 
 def get_module(name: str) -> object:
-    """Return the loaded kernel module for ``name``, or ``None`` if unavailable."""
-    return _modules.get(name)
+    """Return the loaded kernel module for ``name``, importing it on first use.
+
+    Raises ``RuntimeError`` if the kernel is unavailable (not built, or failed
+    to import) — callers that can tolerate a torch fallback should check
+    ``is_available(name)`` first instead.
+    """
+    mod = _try_load(name)
+    if mod is None:
+        raise RuntimeError(
+            f"CUDA kernel '{name}' is not available. "
+            f"Build with CSRC_KERNELS=true (or use the torch-native fallback)."
+        )
+    return mod