Refactor pipeline parallel multi-stage implementation

[plusbang]

Description

Background: https://github.com/analytics-zoo/nano/issues/1405

Support multi-stage PP for both inference and serving, and solve installation problems.

TODO: Has verified multi-stage PP serving locally and will update it later.

2. User API changes

from ipex_llm.transformers import AutoModelForCausalLM, init_pipeline_parallel

init_pipeline_parallel()

model = AutoModelForCausalLM.from_pretrained(model_path,
                                             load_in_4bit=True,
                                             optimize_model=True,
                                             trust_remote_code=True,
                                             use_cache=True,
                                             pipeline_parallel_stages=2)

3. Summary of the change

• Replace model layers with dummy layer in from_pretrained and add pipeline_parallel_generate to pretrained model dynamically
• Update PP inference example

4. How to test?

• Unit test
• Local test

[jason-dai]

Please document which models have been verified.

[plusbang]

Please document which models have been verified.

Have updated verified model list.

[jason-dai]

See comments inline; otherwise LGTM

[jason-dai]

LGTM