At least one test fails on x86_64. GitHub indicates that there are
probably other test failures on other arches.

https://github.com/NixOS/patchelf/issues?utf8=%E2%9C%93&q=is%3Aissue+is%3Aopen+set-rpath-library.sh

Ruby pretends editline (libedit) is readline.  This is normally fine.
However, editline's readline emulation does not account for changes to
rl_instream or rl_outstream.  If a Ruby application using the Readline
extension changes .input or .output, this will cause a use-after-free:

==32694== Invalid read of size 4
==32694==    at 0x4070A38: fwrite (fwrite.c:32)
==32694==    by 0x406F707: fputs (fputs.c:7)
==32694==    by 0x660EAC7: el_wgets (in /usr/lib/libedit.so.0.0.60)
==32694==    by 0x6607BDB: el_gets (in /usr/lib/libedit.so.0.0.60)
==32694==    by 0x6620EDB: readline (in /usr/lib/libedit.so.0.0.60)
==32694==    by 0x65DE3D3: readline_get (readline.c:346)
==32694==    by 0x4DFA49B: rb_protect (eval.c:996)
==32694==    by 0x65DE4FF: readline_readline (readline.c:507)
==32694==    by 0x4F7AD4B: call_cfunc_m1 (vm_insnhelper.c:1723)
==32694==  Address 0x67503dc is 140 bytes inside a block of size 1,264 free'd
==32694==    at 0x490AFC0: free (in /usr/lib/valgrind/vgpreload_memcheck-ppc64be-linux.so)
==32694==    by 0x406D44F: fclose (fclose.c:35)
==32694==    by 0x65DE013: clear_rl_outstream (readline.c:365)
==32694==    by 0x65DE0DF: readline_s_set_output (readline.c:599)
==32694==  Block was alloc'd at
==32694==    at 0x49092C0: malloc (in /usr/lib/valgrind/vgpreload_memcheck-ppc64be-linux.so)
==32694==    by 0x406BEC7: fdopen (__fdopen.c:21)
==32694==    by 0x65DE103: readline_s_set_output (readline.c:603)

Since rl_instream and rl_outstream are read on each rl_initialize, and
editline's rl_initialize is smart enough to not leak memory if it is
called multiple times during program execution, we use this as a way to
force re-reading of rl_instream and rl_outstream.

I hate this patch; I really do.  Better fixes are highly welcome.

system/kbd: fix some regressions in newer releases

See merge request !327

Actually, fix tmux instead of xterm

See merge request !326

Since kbd=2.0.4, three regressions have been discovered:

* CFLAGS were being overriden.
  Fixed by https://github.com/legionus/kbd/commit/15a74479f904f6b15f31854455656710e9aa1942

* loadkeys -d would fail if defkeymap.map.gz was a symlink.
  Fixed by https://github.com/legionus/kbd/commit/acf93e44f6d036303f95555069031f6fb12ce9d1

* setfont would fail if its argument was a compressed font file.
  Fixed by https://github.com/legionus/kbd/commit/7e27102b6fc6991a6a4eca422b513781a26b1639

Since the last commit introduced new binary files for the test suite,
use a handrolled tarball with the changes instead of patching.

user/xterm: fix regression when using tmux horizontal splits

See merge request !325