afmf-linux v1.1.0

Fixed

• GPU hang on Intel (ANV): the block search's cross-subgroup sums and minimums assumed two
  subgroups of 32 lanes, so on the 8- or 16-lane subgroups ANV picks the early-outs differed
  between subgroups and the group split at its next barrier. They now combine however many
  subgroups the driver chose, through shared memory indexed by subgroup id; the same on RADV.
• The scene change detector's divergence pass runs before the coarsest search again instead
  of alongside it: that search returns early on a cut, and reading the verdict while it was
  being written could split a group the same way.
• The scene change detector's shared-memory reductions ran their last six steps without
  barriers, which only holds when the first 32 lanes are one subgroup; on Intel's 8- and
  16-lane subgroups a lane could read a partial sum before it was written, so the detector's
  value was wrong. Every step now has a barrier.
• Generation checks what the flow shaders need before starting instead of failing inside them:
  subgroup basic, arithmetic and quad operations in compute, and storage support for the
  r8ui luma and rg16i flow images; a device without them gets a log line and repeated
  frames. The device line at debug level shows the subgroup size and operations.