feat: unify attention backend with multi-dim mask support

- Add attention() functional entry delegating to active backend
- GQA/MLA forward calls attention() instead of inline cache/SDPA
- CUDA kernels support 2D/3D/4D mask via mask_h_stride field
- CudaBackend.fwd_decode builds 2D padding mask for mixed seq_lens
- KVCache.max_len precomputed in bind_tasks to avoid GPU sync
- batch==1 decode short-circuits mask=None
- Split tests into conftest, test_backend, test_backend_equivalence, test_kernel_mask
- 440 tests pass, L20 decode 1.44-1.60x speedup vs torch native

tests/extension/test_backend.py

@@ -0,0 +1,45 @@
+"""Backend selection and context-manager switching tests.
+
+These tests do not require CUDA — they only check that the active
+backend is correctly set and restored.
+"""
+
+import pytest
+
+from astrai.extension import (
+    ATTN_BACKEND,
+    CudaBackend,
+    TorchNativeBackend,
+    attn_backend,
+    get_backend,
+)
+
+
+def test_default_backend_is_torch_native():
+    backend = get_backend()
+    assert isinstance(backend, TorchNativeBackend)
+
+
+def test_attn_backend_context_with_enum():
+    with attn_backend(ATTN_BACKEND.CUDA):
+        assert isinstance(get_backend(), CudaBackend)
+    assert isinstance(get_backend(), TorchNativeBackend)
+
+
+def test_attn_backend_context_with_class():
+    with attn_backend(CudaBackend):
+        assert isinstance(get_backend(), CudaBackend)
+    assert isinstance(get_backend(), TorchNativeBackend)
+
+
+def test_attn_backend_context_with_instance():
+    custom = CudaBackend()
+    with attn_backend(custom):
+        assert get_backend() is custom
+    assert isinstance(get_backend(), TorchNativeBackend)
+
+
+def test_cudabackend_is_context_manager():
+    with CudaBackend():
+        assert isinstance(get_backend(), CudaBackend)
+    assert isinstance(get_backend(), TorchNativeBackend)