perf(moe): translate CCCL smem_resource_raw.cuh — buffer reuse for MoE output

smem_resource_raw.cuh entire design (180 lines):
  Manages shared memory as multi-stage pipeline resources.
  Core idea: one memory region, multiple stages, barrier-synchronized.
  - mStageCount stages share the same SMEM base pointer
  - data() returns mPtrBase + mStageCurrent * mStride (stage rotation)
  - incrementStage() rotates, parity flips on wraparound
  - release/acquire protocol for producer-consumer sync
  Key insight: allocate once, reuse forever via stage rotation + zeroing.

Translation to MoE _pure_pytorch_experts:
  Previous: torch.zeros_like(hidden_states) every call — GPU malloc + memset.
  Now: class-level _moe_out_buf, resized only when shape changes, .zero_()
  in-place (memset only, no malloc). On BI-V100 without async allocator,
  this eliminates a synchronous cudaMalloc per MoE layer per forward pass.
  With 28 MoE layers × 2 calls/step (prefill+decode), that is 56 fewer
  allocations per step.

CCCL source: cub/cub/detail/warpspeed/resource/smem_resource_raw.cuh
Maps to: qwen3_6_scripts/qwen3_5.py (Qwen3_5MoeSparseBlock)

qwen3_6_scripts/qwen3_5.py

@@ -988,7 +988,17 @@ class Qwen3_5MoeSparseBlock(nn.Module):
             # to enable batched GEMM across expert groups (CCCL segmented_reduce pattern).
             # TODO: implement when we have benchmark data showing this path is hot.
 
-            out = torch.zeros_like(hidden_states)
+            # smem_resource_raw.cuh: reuse buffer across calls.
+            # CCCL manages SMEM as multi-stage ping-pong: same memory, different
+            # stages. We do the same: keep a class-level buffer, resize only if
+            # shape changes, zero in-place instead of allocating.
+            _buf_key = (T, hidden_states.shape[-1])
+            if not hasattr(self, '_moe_out_buf') or self._moe_out_buf_key != _buf_key:
+                self._moe_out_buf = torch.zeros_like(hidden_states)
+                self._moe_out_buf_key = _buf_key
+            else:
+                self._moe_out_buf.zero_()
+            out = self._moe_out_buf
 
             # Flatten all (token, expert) assignments: (T*top_k,) pairs
             flat_eids = topk_ids.view(-1)                      # (T*K,)