fix: pass skip_softmax_threshold_scale_factor to prefill wrapper in test

The wrapper consistency check in _test_trtllm_batch_prefill was calling
wrapper_trtllm_gen.run() without skip_softmax_threshold_scale_factor,
causing it to default to None (standard attention kernel) while the raw
API used 1e-30 (skipsSoftmax kernel variant). Different cubin kernels
produce bit-different results, failing the exact-equality assert.

The decode counterpart was already fixed; this mirrors that fix for the
prefill test path.

Author: PerkzZheng

tests/attention/test_trtllm_gen_attention.py

@@ -849,6 +849,7 @@ def _test_trtllm_batch_prefill(
             v_scale=v_scale / o_scale,
             enable_pdl=enable_pdl,
             sinks=(sink if enable_sink else None),
+            skip_softmax_threshold_scale_factor=skip_softmax_threshold_scale_factor,
         )
         # v_scale, o_scale in wrapper is emulated by multiplying output by v_scale instead of fused into kernel.
         if v_scale == o_scale == 1.0: