Update pipeline parallel serving for more model support

[plusbang]

Description

As a follow-up to #11319:

• Move ModelRunner (implementation for pipeline parallel multi-stage serving) to source code
• Update ModelRunner.load_model for different model support
• Align ModelRunner.model_step with pipeline_parallel_generate about model forward and past key values
• Some small fixes about serving example (remove unused function, add license)

1. Why the change?

Align PP serving and inference, therefore verified models in PP inference could be directly used in serving.

2. User API changes

from ipex_llm.transformers import ModelRunner

4. How to test?

• Unit test
• Local test (run llama2-7b, baichuan2-13b and chatglm3-6b with pp serving example)

[xiangyuT]

Need to pass attention_mask = make_attention_mask(cur_batch.prompt_lengths).to(input.device) here for batch serving.

[plusbang]

Need to pass attention_mask = make_attention_mask(cur_batch.prompt_lengths).to(input.device) here for batch serving.

Have updated :)

[xiangyuT]

LGTM