text2image: Add pimp approach for unet_2d_condition & batch-size 1 implementation of unet for initial NPU support

[RyanMetcalfeInt8]

For now, NPU can be used like this. The unet model must be reshaped to a static shape before compile is invoked.

std::filesystem::path root_dir = models_path;
auto pipe = ov::genai::Text2ImagePipeline::stable_diffusion(
    ov::genai::Text2ImagePipeline::Scheduler::from_config(root_dir / "scheduler/scheduler_config.json"),
    ov::genai::CLIPTextModel(root_dir / "text_encoder", "CPU"),
    ov::genai::UNet2DConditionModel(root_dir / "unet")
        .reshape(2, 512, 512, 77)
        .compile("NPU", ov::cache_dir("./cache")),
    ov::genai::AutoencoderKL(root_dir / "vae_decoder", "GPU", ov::cache_dir("./cache")));
ov::Tensor image = pipe.generate(prompt);

[RyanMetcalfeInt8]

Rebased against latest master branch (resolved text2image -> image_generation move, etc.). @TolyaTalamanov, @ilya-lavrenov, can someone help review?