[winding-lines]

Since 03aac93 ("keep the OpenSSL shim free of thread-local storage"), every fork-based test in tests/tls/test_tls.mojo segfaults on macOS. The first is test_tls_connect_valid_cert_succeeds, and the process exits with 139.

Cause:

• dyld never unloads an image that has thread-local variables, so the old thread_local error buffer was keeping libflare_tls mapped.
• With the buffer now a pthread key, the last dlclose (when _TlsTestServer / TlsStream drop their OwnedDLHandle) really unmaps the shim.
• OpenSSL and the key's destructor still hold pointers into it, so the next call into those addresses crashes.
• Linux avoids this because build.sh links with -z nodelete, which macOS's linker doesn't have.

Evidence, on 8fe83c5e, Mojo 1.2.0, osx-arm64:

• Before: test_tls.mojo crashes in libKGENCompilerRTShared right after the first test's server handle drops.
• Holding one extra dlopen(libflare_tls) for the whole run: 17/17 pass, which isolates the unload as the cause.
• With this change: 17/17 pass.

Fix:

• A constructor in openssl_wrapper.cpp takes an RTLD_NOLOAD | RTLD_NODELETE reference to its own image, found through dladdr. That's the portable spelling of -z nodelete.
• It also covers recipe.yaml's build, which never passed -z nodelete, so the conda package was unloadable on Linux too.
• -ldl is added to both builds for older glibc sysroots.
• No TLS segment is reintroduced, so 03aac93's __tls_get_addr win is kept.

[ehsanmok]

Great! Thanks.