add initial support for intel npu acceleration library

[MeouSker77]

Description

Add initial support for intel npu acceleration library, just keep ipex-llm's API, calling intel_npu_acceleration_library's compile function directly

2. User API changes

from ipex_llm.transformers.npu_model import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, load_in_low_bit='fp16')

ipex_llm.transformers.npu_model.AutoModelForCausalLM receives all arguments of ipex_llm.transformers.AutoModelForCausalLM, but it doesn't support most of them, so it just ignores them.

load_in_low_bit only supports sym_int4, sym_int8 and fp32

Note: import ipex will cause intel_npu_acceleration_library cannot find NPU, so we must disable ipex auto-import when use it by setting set BIGDL_IMPORT_IPEX=0

[rnwang04]

how about fp16?

[MeouSker77]

it only moves Linear Computation to NPU, other OPs are still run on CPU, CPU doesn't support most fp16 operations

[rnwang04]

others LGTM as an initial PR.

[jason-dai]

Note: import ipex will cause intel_npu_acceleration_library cannot find NPU, so we must disable ipex auto-import when use it by setting set BIGDL_IMPORT_IPEX=0

Maybe add another ENV "BIGDL_USE_NPU" to disable auto import

[MeouSker77]

Maybe add another ENV "BIGDL_USE_NPU" to disable auto import

added it

[jason-dai]

LGTM