Crash fix in case of forks: use static TLS for alloc tracking storage (…

…DataDog#490)

* Remove pthread_key and use initial-exec TLS instead

* Replace heap-allocated void* TLS with a char[] buffer using
initial-exec __thread and placement new.

- Add `initialized` field to TrackerThreadLocalState
- Shared C header (tls_state_storage.h) for size/alignment
  constants between loader.c and C++ code, with static_assert
- Clean up dead code: _key_once, _tl_state_key, delete_tl_state
- Add standalone fork test validating TLS zero-init, fork
  inheritance, and fresh-thread behavior

---------

Co-authored-by: Nicolas Savoire <nicolas.savoire@datadoghq.com>