ggml-vulkan: serialize racy vk_instance initialization

[d-e-s-o]

When calling whisper_init_from_file_with_params_no_state() from multiple threads with the Vulkan backend enabled, segfaults abound. The problem seems to be that vk_instance initialization is not serialized properly. Introduce a mutex to fix the problem. It needs to be recursive because of the following call chain:

  ggml_vk_get_device()
  -> ggml_backend_vk_reg()
  -> ggml_vk_instance_init()

[ggerganov]

Thanks for reporting. Could you provide a minimal example of how you use libwhisper in order to trigger this race condition? Ideally a self-contained program that reproduces the segfault.

[d-e-s-o]

I am using it via the whisper-rs Rust bindings, so that's all I can offer.

• check out https://github.com/d-e-s-o/whisper.cpp/tree/topic/repro
• run cd repro
• run cargo test
  • watch nasal daemons emerge

[HDJohnbot]

PM note: this PR is blocked on reproduction details. Maintainer asked for a minimal self-contained repro, and the author’s current pointer is the Rust bindings path rather than a standalone case. Until a minimal repro lands, this should stay blocked rather than drifting. — little John