Test_parseTCPStats fails on big-endian archs

[dswarbrick]

Test_parseTCPStats fails on s390x and presumably other big-endian archs (e.g. mips, mips64, ppc64):

=== RUN   Test_parseTCPStats
    tcpstat_linux_test.go:80: want tcpstat number of bytes in tx queue 42, got 704643072
    tcpstat_linux_test.go:83: want tcpstat number of bytes in rx queue 22, got 369098752
--- FAIL: Test_parseTCPStats (0.00s)

This is due to a wrong assumption about the endianness of s390x etc. in the test:

func Test_parseTCPStats(t *testing.T) {
    encode := func(m InetDiagMsg) []byte {
        var buf bytes.Buffer
        err := binary.Write(&buf, binary.LittleEndian, m)
        if err != nil {
            panic(err)
        }   
        return buf.Bytes()
    }
...

Most, but not all modern CPUs are little-endian. s390x is big-endian.

This is a regression in v1.4.0 due to #2322. Other functionality related to that PR may also be buggy on s390x, in addition to the failing test.

[dswarbrick]

A possible solution might be to use github.com/josharian/native to provide the system's native endianness at compile time, as mentioned in the github.com/mdlayher/netlink changelog:

## v1.2.1

...
- [Improvement]
  [commit](https://github.com/mdlayher/netlink/commit/8014f9a7dbf4fd7b84a1783dd7b470db9113ff36):
  the library now uses https://github.com/josharian/native to provide the 
  system's native endianness at compile time, rather than re-computing it many
  times at runtime.