Treat riscv64 like mips64el by ignoring the border-one-rounded gsk reftest

…and accepting slightly different results for opacity & opacity-overdraw reftests

@smcv The failure output of https://buildd.debian.org/status/fetch.php?pkg=gtk4&arch=riscv64&ver=4.12.1%2Bds-2&stamp=1693193191 looks almost exactly like https://buildd.debian.org/status/fetch.php?pkg=gtk4&arch=mips64el&ver=4.12.0%2Bds-2&stamp=1692729364

The pixel difference count matches and the base64 diff at the bottom of the build logs matches.

Therefore, I think riscv64 needs the same workaround.