[Attention] Use sparse prefill kernel for fp8 kv-cache in DeepSeek-v3.2 (#27532)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

tests/kernels/moe/test_batched_moe.py

@@ -248,6 +248,7 @@ def test_fused_moe_batched_experts(
     per_act_token_quant: bool,
     block_shape: list[int] | None,
     input_scales: bool,
+    workspace_init,
 ):
     """Note: float8_e4m3fn is not supported on CUDA architecture < 89,
     and those tests will be skipped on unsupported hardware."""