fix: 384 experts (not 256), clarify cross-rank reduce is in caller

DeepSeek-V4-Pro has 384 routed experts, 48 per rank (384/8).
The cross-rank all-reduce happens in the parent DeepseekV4MoE.forward,
not in our kernel. Our kernel writes local output; caller does reduce.
Fixed README, nvfp4_mega_moe.py comments.

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -7,7 +7,7 @@ Architecture:
 - L1 GEMM: gate_up_proj (FP4 x FP4 → BF16 with UE4M3 scales)
 - SiLU+Mul activation
 - L2 GEMM: down_proj (FP4 x FP4 → BF16 with UE4M3 scales)
-- NVLink cross-rank sync via symm buffer
+- NVLink cross-rank sync handled by caller (not this kernel)
 - Expert parallel: each rank handles NUM_EXPERTS/8 experts
 
 The kernel uses native NVFP4 block-scaled MMA via tcgen05.mma
@@ -253,7 +253,7 @@ def nvfp4_mega_moe_full(
     3. SiLU + Mul (activation)
     4. Quantize L1 output → FP4 + UE4M3 scales
     5. L2 GEMM: down_proj (native NVFP4 block-scaled MMA)
-    6. NVLink sync + reduce across ranks → write to y
+    6. Write to y (caller handles cross-rank all-reduce)
 
     Uses tcgen05.mma.kind::mxf8f6f4.block_scale for native E2M1×E2M1
     with UE4M3 block-16 scaling in Blackwell tensor cores.
@@ -305,5 +305,5 @@ def nvfp4_mega_moe_full(
         topk_ids, topk_weights, num_experts_per_rank,
     )
 
-    # Step 6: Write to output
+    # Step 6: Write to output (caller handles cross-rank all-reduce)
     y.copy_(l2_output)