This changes how Scheduler::Flush works. It queues the current command
buffer to be sent to the GPU but does not do it immediately. The Vulkan
worker thread takes care of that. Users will have to use
Scheduler::Flush + Scheduler::WaitWorker to get the previous behavior.
Scheduler::Finish is unchanged.
To avoid waiting on work never queued, Scheduler::Wait sends the current
command buffer if that's what the caller wants to wait.
Move code to separate files to be able to reuse it from OpenGL. This
greatly simplifies the pipeline cache logic on Vulkan.
Transform feedback state is not yet abstracted and it's still
intrusively stored inside vk_pipeline_cache. It will be moved when
needed on OpenGL.
Move descriptor lookup and update code to a separate thread. Delaying
this removes work from the main GPU thread and allows creating
descriptor layouts on another thread. This reduces a bit the workload
of the main thread when new pipelines are encountered.
Create multiple descriptor pools on demand. There are some degrees of
freedom what is considered a compatible pool to avoid wasting large
pools on small descriptors.
Avoid using std::array to fix Intellisense not properly compiling this
code and disabling itself on all files that include it.
While we are at it, change the code to use u8 instead of size_t for the
number of instructions in an opcode.
When we can't track the SSBO origin of a global memory instruction,
leave it as a global memory operation and assume these pointers are in
the NVN storage buffer slots, then apply a linear search in the shader's
runtime.
