Give the Vulkan allocator one owning thread

Readback buffers are allocated when the readback is recorded rather than on the render thread,
so only the thread that records frames touches VMA. The allocator is created externally
synchronized, and VMA is built with no-op mutexes and plain counters instead of std::mutex and
std::atomic.

Headless runs Vulkan offscreen in builds without SDL too, and CI runs the headless tests on
lavapipe, skipping the pixel tests a hardware backend doesn't match the software references on.
sleep_ms and sleep_us go, with the thread manager test waiting on an event instead, and so does
the unused portable-file-dialogs header.

.github/workflows/build.yml

@@ -345,7 +345,7 @@ jobs:
       if: runner.os == 'Linux'
       run: |
         sudo apt-get update -y -qq
-        sudo apt-get install libsdl3-dev libgl1-mesa-dev libglu1-mesa-dev libsdl3-ttf-dev libfontconfig1-dev libcurl4-openssl-dev
+        sudo apt-get install libsdl3-dev libgl1-mesa-dev libglu1-mesa-dev libsdl3-ttf-dev libfontconfig1-dev libcurl4-openssl-dev mesa-vulkan-drivers
 
     - name: Install macOS dependencies
       if: runner.os == 'macOS'
@@ -390,6 +390,12 @@ jobs:
         done
         exit $fail
 
+    # Runs the Vulkan backend's render and compile threads on lavapipe.
+    - name: Execute headless tests (Vulkan)
+      if: matrix.os == 'ubuntu-26.04'
+      working-directory: ${{ env.GITHUB_WORKSPACE }}
+      run: python3 test.py -g --graphics=vulkan --known-failures=vulkan
+
     # x86-64 only: the reference images don't match what the software renderer produces on arm64,
     # 23 of the 30 dumps differ there, one of them by an MSE of 100. That's a real difference and
     # worth chasing, but it predates the arm64 runner - it just never had a CI job to show up in.