[Feature]Moe alltoallv communication optimization for unquantized RL training sence & alltoallv support dpo

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>

Author: weijinqian_v1

vllm_ascend/ops/fused_moe.py

@@ -991,7 +991,7 @@ def apply(
                 global_batch_size=self.global_batch_size,
                 expert_map=expert_map,
                 ep_group=get_ep_group())
-        elif fused_moe_state == FusedMoEState.All2AllSeq is not None and is_prefill:
+        elif fused_moe_state == FusedMoEState.All2AllSeq and is_prefill:
             token_dispatcher = kwargs.get('token_dispatcher')
             return fused_experts_with_all2allv(token_dispatcher=token_dispatcher,
                                                probs=topk_weights,