Update DT_VERDEF when .gnu.version_d is relocated

[Wenri]

Fixes #664.

The bug

rewriteHeaders() writes the new addresses of relocated sections back into
.dynamic, but only for DT_VERSYM and DT_VERNEED. DT_VERDEF has no
branch, so whenever .gnu.version_d is moved the tag keeps pointing at the
pre-relocation address. glibc then follows a stale verdef chain in
_dl_check_map_versions, which can segfault before main() runs.

The fix mirrors the branch that already exists for .gnu.version_r:

else if (d_tag == DT_VERDEF)
    dyn->d_un.d_ptr = findSectionHeader(".gnu.version_d").sh_addr;

Why it has gone unnoticed

Two things have to line up, and the existing tests never line them both up:

1. The object must define symbol versions, not merely require them.
   Every current test patches objects that have .gnu.version_r (already
   handled); none patches one with .gnu.version_d.
2. The linker's section ordering must put .gnu.version_d in the way.
   lld places .dynstr after the .gnu.version* sections, so growing
   .dynstr drags .gnu.version_d along. GNU ld and gold place it before,
   and there the section is never relocated and the bug is invisible.

It is not new — I reproduced it on 0.15.2, 0.16.1, 0.17.2, 0.18.0 and
0.19.1 (master 7688b17). Whether a stale tag actually crashes depends on
what happens to sit at the old address, so it can also ship silently as a
corrupt-but-running binary; that is the more unpleasant half.

Regression test

tests/verdef-relocate.sh builds a non-PIE executable exporting a versioned
symbol (-rdynamic plus a version script), applies an rpath too long to edit
in place so the dynamic cluster is relocated, and asserts DT_VERDEF still
points at .gnu.version_d. It also runs the patched binary.

The address comparison is the real oracle rather than "does it still run",
because surviving a stale tag is luck.

The linker is chosen by probing for one whose layout actually moves the
section, not by name, so the test starts exercising the bug automatically
if a default toolchain changes ordering — and exit 77 (skip) if none does,
rather than passing vacuously.

Verified:

	result
upstream 0.19.1, 0.15.2	FAIL: DT_VERDEF is '0x2003d8' but .gnu.version_d moved to 0x1fe410
with this patch	PASS
no suitable linker	SKIP (77)

make check on this branch: 70 tests, 66 PASS, 4 SKIP, 0 FAIL.
shellcheck tests/*.sh is clean.