Vulkan: ggml_backend_vk_set_tensor_2d_async() performance tuning

[UltraRabbit]

I could observe there would be a 6 ~ 7 seconds cost to the first call of this function, but the following calls of this function just hold for 0.001 seconds. I guess there might be some resources to be initialized at the first place. This would happen for each round of user request, which always introduce a 6 ~ 7 seconds delay for the first response to user. Is it possible to avoid realloc those resources repeatly at every round of user request?

[sanjayahari]

@UltraRabbit I looked at ggml_vk_ensure_sync_staging_buffer() and it seems the staging buffer is only allocated on first use would pre-allocating it in the init path help?