LLM: split chatglm3's mlp and use mlp fusion

[rnwang04]

Description

This split chatglm3's mlp and use mlp fusion, which can has ~1ms on MTL. But quantize kv cache + mlp fusion will cause change of output on Arc & MTL (which seems a known issue)

1. Why the change?

2. User API changes

3. Summary of the change

4. How to test?

• [ ] Unit test