P5: add reference comparison to Python multi-tile test

tests/unit/test_p5_python_minimal.py

@@ -20,6 +20,16 @@ try:
     o, lse = fmha_multitile_decode_raw(q, k, v, scale)
     print(f'Output[0,0,0,:5]: {o[0,0,0,:5].float()}')
     print(f'LSE: {lse[0,0,0].item():.4f}')
-    print('SUCCESS')
+
+    # Reference
+    q_r = q[0,0].float()  # (1, hd)
+    k_r = k[0,0].float()  # (N, hd)
+    v_r = v[0,0].float().T  # (N, hd) — V is (hd, N), transpose for reference
+    s = torch.matmul(q_r, k_r.T) * scale
+    s = torch.softmax(s, dim=-1)
+    o_ref = torch.matmul(s, v_r)
+    cos = torch.nn.functional.cosine_similarity(o[0,0].float().flatten().unsqueeze(0), o_ref.flatten().unsqueeze(0)).item()
+    print(f'Cosine vs reference: {cos:.6f}')
+    print(f'{"PASS" if cos >= 0.999990 else "FAIL"}')
 except Exception as e:
     print(f'FAILED: {e}')