try fit in new transformers moe

[ETOgaosion]

transformers>5.0 changes All MoE weights format, now there is a large expert gate_up tensor and down tensor:

https://github.com/huggingface/transformers/blob/aad13b87ed59f2afcfaebc985f403301887a35fc/src/transformers/models/qwen3_5_moe/modeling_qwen3_5_moe.py#L820-L821

        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))

[ETOgaosion]

Now we support: loading legacy model save checkpoints with experts.x.gate_up_proj (according to actual key, also support new API)

And saving models with new state_dict() API with experts.gate_up_proj (according to transformers_version in transformer configs, also support legacy API)