fix: use full padded_scales_buf (no GPU scalar slicing in cudagraph)

buf[:gpu_scalar, :] triggers cudaErrorStreamCaptureInvalidated.
Always use the full pre-allocated buffer; extra rows are zeros.

vllm/nvfp4_cutedsl.py

@@ -162,10 +162,8 @@ class CuTeDSLMoERunner:
         padded_expert_offsets.zero_()
         padded_expert_offsets[1:] = padded_rows_per_expert.cumsum(0)
         
-        total_padded_rows = padded_expert_offsets[-1]
-        
-        # Reset the padded scales buffer
-        padded_scales = self._padded_scales_buf[:total_padded_rows, :padded_cols]
+        # Use the FULL pre-allocated scales buffer (no GPU scalar slicing)
+        padded_scales = self._padded_scales_buf
         padded_scales.zero_()
         
         # Build index mapping: for each row in x_sf, which expert does it belong to?