afmf-linux v0.3.0

First public release.

Added

• Presentation thread per swapchain: the application's thread returns right after the layer's
  submit; both presents, the spare refill and the pacing hold run on the thread.
• Half-frame pacing (AFMF_PACING, on by default): the real frame is held back half the
  smoothed frame time so the generated frame lands evenly in between, the same half-frame of added
  latency AMD's implementation has.
• Shared compute queue: when the application took every compute queue (vkd3d-proton asks for all
  of RADV's), the layer works on the application's last one and serialises the application's own
  submits on it.
• Spare image acquired a frame ahead with a fence, never waited for in the present hook
  (AFMF_ACQUIRE_TIMEOUT_US defaults to 0).
• Host-side profile (AFMF_PROFILE=1): real fps, microseconds spent in the hook and on the
  thread, pacing hold; surfaces log their window system (Wayland, X11, headless).
• vkAcquireNextImage(2)KHR and vkDeviceWaitIdle hooks, serialised with the thread.
• Present chains carried to the thread: VkPresentIdKHR, VkSwapchainPresentModeInfoEXT,
  VkSwapchainPresentFenceInfoEXT, VkPresentTimingsInfoEXT; anything else presents inline.
• AFMF_SEARCH_MODE=auto picks 5 pyramid levels at half flow resolution, 7 at full.
• Headless test variants: shared queue (AFMF_TEST_ALL_QUEUES), paced presents
  (AFMF_TEST_FRAME_MS, AFMF_TEST_FRAMES).

Changed

• Barriers between compute passes no longer name the transfer stage; the pyramid and the scene
  change histogram run without a barrier between them (493 -> 434 us per frame at 3440x1440).