Very simple algorithm that just naively mixes in some of the pre-tone mapped high frequency I channel into the tone-mapped output. Using a kernel size of 8x8 gives fairly good results here.
This commit also bundles a logic change in pl_renderer to decide whether or not to enable peak detection - as a result of contrast recovery, we want to do our own peak detection even if we already have metadata. The only case this substantially alters is the question of whether or not static HDR10+ scene metadata is better than dynamic DV PQ metadata, and the addition of CR info arguably tips that balance.