path: serialize iconv_open and iconv_close - #7404

Fixes #7403.

On macOS 14 and later, iconv_open and iconv_close take a process-wide pthread rwlock that can lose a wakeup when a waiting thread takes a signal. Once that happens, every thread that opens or closes a converter blocks forever. We open and close a converter for every directory we iterate and every reference name we normalize whenever core.precomposeunicode is set, which is the default on APFS. So a multithreaded program that handles any signal can hang in a reference lookup, a status or a diff. #7403 has the details, the Apple source references and a reproducer.

This change holds a process-wide git_mutex around iconv_open and iconv_close. The mutex is created by git_fs_path_iconv_global_init, which git_libgit2_init runs alongside the other util-layer inits; it is freed through git_runtime_shutdown_register. With only one thread ever inside those calls, iconv's own lock never has a waiter. The conversions themselves (iconv()) don't take that lock and stay unserialized.

Notes for review:

• git_fs_path_iconv_init_precompose now sets map to (iconv_t)-1 before it takes the lock, so a failed lock leaves ic in a state git_fs_path_iconv_clear accepts. A lock failure is reported as GIT_ERROR_THREAD, as in mwindow.c.
• The lock call is checked for a non-zero return, as in mwindow.c and openssl.c. On POSIX, git_mutex_lock is pthread_mutex_lock, which returns a positive errno, so a < 0 check would never fire.
• Builds without GIT_I18N_ICONV compile all of this out, including the init entry.
• No test is added. The failure needs a signal to land on a thread that is waiting inside the OS's rwlock. It is timing-dependent and macOS-only, and a regression would hang the suite rather than fail it. The existing iconv, path, refs, iterator and threads tests exercise the lock, and tests/util/init.c exercises its init and shutdown.

Testing, on macOS 26.5.1 with an Apple M3 Ultra:

• Reproducer from Threads hang forever in iconv_open/iconv_close on macOS 14+ when a signal arrives (core.precomposeunicode) #7403 (64 threads, SIGUSR1 every 200 µs). Unpatched main wedged in 3 of 3 runs, and in 5 of 5 runs of a longer harness. Patched main kept going for the full 20 s in 3 of 3 and 5 of 5 runs, and for a 300 s soak: 6.3M lookups under about 985k signals.
• Backport. The same change applied to v1.9.7, with GIT_USE_ICONV in place of GIT_I18N_ICONV, went from 3 of 3 runs wedged to 3 of 3 kept going.
• Test suites. util_tests and libgit2_tests -xonline pass. The util suite and the status::renames, refs, repo::init, iterator and threads suites give identical results with and without the patch, on main and on v1.9.7.
• Build configurations. Builds warning-free with -DENABLE_WERROR=ON, with -DUSE_THREADS=OFF, with both together, and with -DUSE_I18N=OFF.
• Throughput. Single-threaded throughput is unchanged, at about 75k lookups/s either way. With 64 threads on one repository, the patched build does about 22k lookups/s against about 75k unpatched. Unpatched libgit2 with precomposition off, which never calls iconv, also does about 22k/s. The unpatched figure is higher only because those threads spend about 95% of their samples parked on iconv's lock, which throttles their file I/O. The mutex adds no measurable cost.