Diff produces a broken table

[Boulder08]

Diff'ing these two files produces a broken film grain table. The coefficients are b0rked somehow, and if you use the table in aomenc, you get a very interesting grain layer there..

https://drive.google.com/file/d/1V2e8DRlNO_KGfEVfiRWO83dwSV2YthkB/view?usp=drive_link
https://drive.google.com/file/d/1hVUvOne-WRNX2QpYfY8w8XETBjQbtedN/view?usp=drive_link
tng s06e02_grain.txt

The same happens if you use the original source (1080p) and crop the left and right side borders (244 pixels both) off and try working on that without downscaling to 720p. If you crop and add the borders back before encoding the reference and lossless clip, no problems so the problem is somewhere in grav1synth.

I also tried various cropping parameters but none of them worked. Cropping the top and bottom borders has not caused any problems in my tests.

[Boulder08]

Looks like widths 1920 and 1280 work. Others have issues, didn't test higher ones yet.

[Boulder08]

Some tests with various widths:
3840 ok
3600 not ok
3584 ok
2960 not ok
2880 not ok
2560 ok
1920 ok
1480 ok
1460 not ok
1456 not ok
1280 ok