freedreno: Autotune bypass vs GMEM rendering decision

freedreno: Autotune bypass vs GMEM rendering decision

In some cases, like gl_driver2, we have all the characteristics that
make our current simplistic bypass vs GMEM decision pick GMEM (ie. batch
starts with a clear, has blend enabled, has a high draw count, etc),
but each draw touches very few pixels and the per-tile state-change
overhead leaves us CP limited.  We would be better in this case picking
the bypass path.

So use feedback from # of samples-passed in previous render passes to
the same FBO to give us a bit more information to make better choices.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2798
Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9535>