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/conftest.py

@@ -0,0 +1,35 @@
+"""Shared fixtures for extension tests."""
+
+import pytest
+import torch
+
+from astrai.config.model_config import AutoRegressiveLMConfig
+from astrai.extension import is_available
+from astrai.model.transformer import AutoRegressiveLM
+
+CUDA_AVAILABLE = torch.cuda.is_available() and is_available("attn_paged_decode")
+skip_no_cuda = pytest.mark.skipif(
+    not CUDA_AVAILABLE, reason="CUDA not available or kernels not built"
+)
+
+D = 64
+CFG = dict(
+    vocab_size=1000,
+    hidden_size=128,
+    num_attention_heads=2,
+    num_key_value_heads=1,
+    intermediate_size=256,
+    max_position_embeddings=64,
+    num_hidden_layers=2,
+    rms_norm_eps=1e-5,
+    attn_type="gqa",
+    ffn_type="mlp",
+)
+
+
+@pytest.fixture
+def cuda_model():
+    config = AutoRegressiveLMConfig(**CFG)
+    model = AutoRegressiveLM(config).to(device="cuda", dtype=torch.bfloat16)
+    model.eval()
+    return model, config