Remaining work on RCD, LMMSE and PPG demosaicers

[jenshannoschwalm]

1. To avoid border remaining CPU/GPU differences we have to increase the outermost margin to 9 as we did for the border
2. As the RCD-internal border algorithms was basically a PPG variant that code got deduplicated and we have that now all in ppg code.
3. Also some PPG CPU/GPU differences got fixed.

@TurboGit there are further changes in expected output as for RCD the CPU border code was still wrong and in PPG we have subtle differences but those are correct now.

Overall, together with #20639 and #20631 this was the round of demosaicer related finetuning & fixes. checked integration tests with all three and couldn't spot any unexpected/bad results.

Currently we have

1. Quite a number of tests with very large CPU/GPU pixel diff counts
2. a few tests where (so it seems to me) the tiling overlaps are problematic as in 0030 0114 ... Not sure what to make out of those atm. Simply don't know if the reference images are wrong or the new ones. if we update any ref images it might be good to do this without the "reference resourcelevel" conf setting to be sure we don't tile and thus later also check for any tiling problems. (Interestingly the pattern is not reproduced in the CPU/GPU diff images.