GHA: Cut down cross-compilation sysroots

In GHA cross-compilation, we don't need everything from the sysroot, because we use native compilers for the compilation. Therefore, we can cut them even deeper and reclaim significant amount of cache space. This would also make cache eviction less likely, which saves time on recreating the sysroots when that happens.

Current sysroots take about 300M per arch, which quickly multiplies up for multiple arches (we build 5 now) users (ballparking at about 200 active contributors). This amounts to about 1500M caches per job, and 300G in caches across our org.