[Mystodan]

On devices without timeline semaphores, emit_implicit_sync_to_queues() runs with the device
lock held (called from submit_queue() / submit_empty_inner()), but calls
request_legacy_semaphore() -> request_semaphore(), which takes LOCK() again on the same
non-recursive std::mutex and hangs forever.

Hit on NVIDIA Tegra X1 (Shield TV, Android 11, Vulkan 1.1.178, driver 495.00) while running the
PyroWave decoder. Backtrace of the hung thread:

std::mutex::lock()
Vulkan::Device::request_semaphore(VkSemaphoreType, VkSemaphore, bool)
Util::for_each_bit<Device::emit_implicit_sync_to_queues(QueueIndices)::$_1>
Vulkan::Device::submit_queue(...)
Vulkan::Device::submit_queue(...)
Vulkan::Device::flush_frame_nolock()
Vulkan::Device::end_frame_nolock()
Vulkan::Device::next_frame_context()
pyrowave_decoder_decode_gpu_buffer

The fix allocates the binary semaphore directly, as the other _nolock paths do. With it, the
decoder runs correctly on that device (output matches a desktop GPU decode at ~57 dB PSNR).

(Context: that driver also lacks VK_KHR_synchronization2; I'm opening a separate PyroWave issue
about Vulkan 1.1 / sync2 support. This fix is independent of that.)

Themaister/pyrowave#30

[Themaister]

Right, the legacy paths don't get tested much. Fix seems correct.