add a new implementation of thread-local storage

[thestinger]

It should map directly to #[thread_local] on most platforms. Android may need a special slow path since it doesn't have a full implementation of TLS. There are various other bugs blocking an implementation that's as efficient and flexible as C++'s thread_local.

• implement destructor support for thread-local variables #17572: implement destructor support for thread-local variables -> must do memory allocation and register a dynamic TLS entry with the destructor

[alexcrichton]

Added in #19094