Significantly reduce thread abuse for faster model moving

This will move all major gradio calls into the main thread rather than random gradio threads.
This ensures that all torch.module.to() are performed in main thread to completely possible avoid GPU fragments.
In my test now model moving is 0.7 ~ 1.2 seconds faster, which means all 6GB/8GB VRAM users will get 0.7 ~ 1.2 seconds faster per image on SDXL.

modules/initialize.py

@@ -149,24 +149,9 @@ def initialize_rest(*, reload_script_modules=False):
     sd_unet.list_unets()
     startup_timer.record("scripts list_unets")
 
-    def load_model():
-        """
-        Accesses shared.sd_model property to load model.
-        After it's available, if it has been loaded before this access by some extension,
-        its optimization may be None because the list of optimizaers has neet been filled
-        by that time, so we apply optimization again.
-        """
-        from modules import devices
-        devices.torch_npu_set_device()
-
-        shared.sd_model  # noqa: B018
-
-        if sd_hijack.current_optimizer is None:
-            sd_hijack.apply_optimizations()
-
-        devices.first_time_calculation()
-    if not shared.cmd_opts.skip_load_model_at_start:
-        Thread(target=load_model).start()
+    from modules_forge import main_thread
+    import modules.sd_models
+    main_thread.async_run(modules.sd_models.model_data.get_sd_model)
 
     from modules import shared_items
     shared_items.reload_hypernetworks()