perf: tune bf16 gemv and add opt-in fused swiglu

- deepen common-shape BF16 GEMV tuning with warp-row tiling for LLaMA/Qwen2/GPT-NeoX/OPT decode projections
- add fused BF16 up/gate SwiGLU CUDA primitive with ASTRAI_SWIGLU=0/1/auto dispatch
- keep the unfused linear backend as the default path; auto enables no shape until per-architecture checkpoint gates pass
- fall back to the linear/torch chain when kernels are absent, on CPU, in training, or outside supported M/K/dtype shapes
- add gemv/swiglu benchmark scripts, dispatch and parity tests, and kernel documentation

Benchmark: NVIDIA L20 (sm_89), CUDA 12.8, PyTorch 2.11.0+cu128, idle GPU. AstrAI 1B config (24 layers, hidden 1536, vocab 100000), BF16, prompt 128, 32 greedy decode tokens, CUDA graphs enabled, A/B in separate interleaved processes (3 rounds, 8 trials each, medians). Default vs ASTRAI_SWIGLU=1 per generate call: batch 1 134.8->129.1 ms (+4.44%), batch 2 136.2->130.9 ms (+4.06%), batch 4 145.5->140.3 ms (+3.66%). Greedy output identical at batch 1, differs at batch 2/4, so auto stays unfused by default; kernelless fallback verified bit-identical greedy.

tests/extension/test_gemv.py

@@ -43,6 +43,66 @@ def test_bf16_gemv_matches_small_decode_batches(m, n, k):
     torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.5)
 
 
+@skip_no_gemv
+@pytest.mark.parametrize("m", [2, 4])
+@pytest.mark.parametrize(
+    "n,k",
+    [
+        (1024, 4096),
+        (4096, 4096),
+        (11008, 4096),
+        (4096, 11008),
+        (14336, 4096),
+        (4096, 14336),
+        (5120, 5120),
+        (13824, 5120),
+        (5120, 13824),
+        (16384, 4096),
+        (4096, 16384),
+        (512, 3584),
+        (3584, 3584),
+        (18944, 3584),
+        (3584, 18944),
+        (1024, 8192),
+        (8192, 8192),
+        (28672, 8192),
+        (8192, 28672),
+        (2048, 2048),
+        (8192, 2048),
+        (2048, 8192),
+    ],
+)
+def test_bf16_gemv_matches_common_transformer_shapes(m, n, k):
+    torch.manual_seed(2026 + m + n + k)
+    x = torch.randn(m, k, device="cuda", dtype=torch.bfloat16)
+    weight = torch.empty(n, k, device="cuda", dtype=torch.bfloat16)
+    weight.normal_(mean=0.0, std=0.02)
+    actual = bf16_gemv(x, weight)
+    expected = F.linear(x, weight)
+    torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.25)
+
+
+@skip_no_gemv
+@pytest.mark.parametrize(
+    "m,n,k",
+    [
+        (1, 8192, 2048),
+        (8, 4096, 11008),
+        (8, 512, 3584),
+        (8, 1024, 8192),
+        (8, 2048, 8192),
+    ],
+)
+def test_bf16_gemv_matches_half_cta_edge_bands(m, n, k):
+    torch.manual_seed(2026 + m + n + k)
+    x = torch.randn(m, k, device="cuda", dtype=torch.bfloat16)
+    weight = torch.empty(n, k, device="cuda", dtype=torch.bfloat16)
+    weight.normal_(mean=0.0, std=0.02)
+    actual = bf16_gemv(x, weight)
+    expected = F.linear(x, weight)
+    torch.testing.assert_close(actual, expected, rtol=0.02, atol=0.25)
+
+
 @skip_no_gemv
 def test_bf16_gemv_preserves_singleton_batch_and_fuses_bias():
     torch.manual_seed(23)