fix pipe device error and enable user control compile

[DefTruth]

This PR fix pipe device error while using device=cpu in model loading and enable user controllable compile setting. After this PR, users can run HunyuanImage-2.1 on many device with low GPU VRAM (< 48GiB). For example:

model_name = "hunyuanimage-v2.1"
pipe = HunyuanImagePipeline.from_pretrained(
    model_name=model_name,
    torch_dtype="bf16",
    # NOTE: load in CPU first, this will enable HunyuanImage run
    # on many device with low GPU VRAM (<96 GiB):
    # CPU -> GPU VRAM < 96GiB ? -> FP8 weight only on CPU -> GPU
    device="cpu" if GiB() < 96 else "cuda",
    use_compile=False,
)

Please also check another PR for more details: #13

@KimbingNg @yestinl