Enalbe fused softmax/sigmoid + topk path for 1024 experts

Per measuring, the fused path delivers better performance when the number of experts is 1024.
1 token + 1024 experts: average uplift ~3%
64 tokens + 1024 experts: average uplift ~6%
128 tokens + 1024 experts: average uplift ~7%
256 tokens + 1024 experts: average uplift ~45%
Current MoE models do not yet support as many as 1024 experts. However, when customers compare performance at 1024 experts, this optimization can provide better performance metrics.

Signed-off-by: LiJianyu <jianyu.li@intel.com>

Author: JianyuLi01

csrc/moe/topk.cpp

@@ -738,6 +738,10 @@ void topk_gating_kernel_launcher(
       LAUNCH_TOPK(
           512, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2, ScoringFuncParam);
       break;
+    case 1024:
+      LAUNCH_TOPK(
+          1024, WARPS_PER_TB, BYTES_PER_LDG_POWER_OF_2, ScoringFuncParam);
+      break;
     case 192:
       LAUNCH_TOPK(
           192, WARPS_PER_TB, BYTES_PER_LDG_MULTIPLE_64, ScoringFuncParam);