Make GPU duration metrics more accurate for Vulkan

In the Vulkan pipeline, the GPU start time was measured to be when
swapBuffers starts. But the command queue has already been submitted at
this point, which means that the GPU work would already have begun. This
means that it's possible to measure a negative time since for very light
GPU workloads, the GPU fence can fire prior to CPU making it to
swapBuffers().

To compensate, instead measure from after skia completed submitting GPU
commands, until the GPU fence fires. Since it is theoretically possible
for GPU work to complete if the render thread gets descheduled
immediately after submitting the GPU, we also add some clamping to
ensure that the duration from command submission -> completion of GPU
work is nonnegative.

Bug: 230713131
Test: atest android.view.cts.FrameMetricsListenerTest
Change-Id: Ia30b7732eaab71e4e29766f788d5cd94ec63c38a