[Vulkan] Reduce peak memory usage when loading models with ET-VK

[SS-JIA]

Context

Currently when running Llama 3.2 1B/3B on Samsung Galaxy S24, the screen blackout may blackout (Llama 3.2 1B) or the device crash (Llama 3.2 3B) when running Llama 3.2 models on Samsung Galaxy S24.

After some investigation I've determined that this behaviour is related to high peak memory usage when loading the model.

cc @manuelcandales @cbilgin

[SS-JIA]

Progress

• [ET-VK] Allocate memory for weight and activation tensors lazily #13474 (top of stack)