Optimized set_initialized_submodules.

[LagPixelLOL]

Fixed an insanely slow function set_initialized_submodules, when loading DeepSeek V3, since it has so many experts, the amount of state_dict_keys are huge and also there are a lot of named_modules.

loaded_keys = {k.replace(f"{module_name}.", "") for k in state_dict_keys if k.startswith(f"{module_name}.")}

This line, especially k.startswith(f"{module_name}."), gets called at least 1 million times, which is insanely slow, it takes at least 10 minutes runtime.

My optimization reduces the runtime to less than 10 seconds by not iterating through state_dict_keys and deleting module_name from the start for every named_module, instead, it prepends the module_name before the module.state_dict() keys.

cc: @ArthurZucker

[ArthurZucker]

Arf sorry for my super late review. This does make sense, don't know why we were re-computing the set so many time.
Thanks 🤗