Bug on Mesa that affects llama.cpp with the Vulkan backend on Linux

[wbruna]

Just a heads-up:

vkEnumerateInstanceExtensionProperties calls close on invalid bytes

In a nutshell: random file descriptors are being closed during the Vulkan's backend initialization, so sometimes fd 1 gets closed, inhibiting the output of commands like llama-bench.

I was able to work around it by first duplicating the stdout descriptor to a backup, then duplicating it back afterwards. Probably not worth including into ggml, unless a lot of people are affected by this.