implement destructor support for thread-local variables

[thestinger]

C++ thread_local has destructor support built on platform primitives. Rust's #[thread_local] does not since it sits above the LLVM layer. This could be feature gated, but it is an important building block for fast TLS support.

[thestinger]

Modern platforms have explicit support for this thanks to C++11, but there will need to be a slow path for older platforms. For example, OS X provides _tlv_atexit and glibc has __cxa_thread_atexit_impl. Rust needs support for selecting code based on the minimum platform version. The slow path can lock a global mutex on the first access to a TLS variable in a new thread and set up a dynamic TLS entry via an API like pthread_key_create.

The normal accesses will go through the static thread_local but dynamic TLS will take care of running destructors. The nightly Linux and OS X builds would need to use this slow path, but distribution builds will use the fast path. It's not worth trying to detect this at runtime because it will only happen once per TLS variable in a thread.

[alexcrichton]

Added in #19094