[NPU] update save-load API usage

[plusbang]

Description

Update save-load API usage.

• save: For all cases with optimize_model=True, specify save_directory is required during the first time to load model, ckpt is saved to save_directory after converting.
• load: still use load_low_bit API to load, and add python cpp backend support

2. User API changes

# first time load, and ckpt is saved to `save_directory` after converting
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    trust_remote_code=True,
    attn_implementation="eager",
    optimize_model=True,
    max_context_len=1024,
    max_prompt_len=960,
    mixed_precision=True,
    quantization_group_size=0,
    save_directory=save_directory  # required and has related check
)

# load converted model
model = AutoModelForCausalLM.load_low_bit(
    save_directory,
    attn_implementation="eager",
    torch_dtype=torch.float16,
    optimize_model=True,
    max_context_len=1024,
    max_prompt_len=960,
    trust_remote_code=True,
)

example and benchmark scripts have updated.

4. How to test?

• Application test

[rnwang04]

LGTM

[plusbang]

Merge it first and will add tokenizer save-load processing in examples later.