Add 3rd input sets to test multi-turn conversation to LLMs

[titaiwangms]

Ideally, we want the exported LLMs pass 3 tests:

1. prompt-processing
2. token generation
3. multi-turn conversation