ENOTSUPP is not included in x/sys/unix since it's meant to be kernel
internal. Move it from our wrapper package into internal/sys.
Also provide a descriptive error message.

Linux on arm64 returns ENOTSUPP when trying to create a tracing link
via BPF_LINK_CREATE and BPF_RAW_TRACEPOINT_OPEN. Turn this into ErrNotSupported.

Use symbols and tracepoints which are portable across amd64 and arm64.

Based on ubuntu 5.13.0-52-generic aarch64.

Don't return unwrapped sentinel errors from various feature tests.

testutils.SkipIfNotSupported will bail out if it encounters an unwrapped
ErrNotSupported.

Trying to create a perf_event uprobe on arm64 doesn't seem to allow
passing an address in the first page. Use the first byte of the second
page instead.

Passing flags to BPF_MAP_LOOKUP_AND_DELETE_ELEM is only possible
from 5.14. Skip the test on earlier kernels.

TestKprobeOffset used to hardcode specific offsets to test against,
which meant we had to make it amd64 specific. It turns out that even
on a single arch the offsets aren't stable due to changes in the
compiler. Hence the test was changed to brute force a valid offset,
but we forgot to run it on all arches.

Trying to execute the test on Linux 5.13 on arm64 triggers EINVAL
instead of the expected EILSEQ however. This is because arm64 returns
EINVAL when trying to use an offset that is not a multiple of four.

Relying on EILSEQ is therefore not portable, and probably shouldn't
be encouraged. Drop the check for EILSEQ from TestKprobeOffset and
execute it on all arches.