This has several implications:
* A race-condition that could manifest itself both on macOS and Linux is
removed. The race could happen if we took the lock to the thread name
mapping while another suspended thread was in the memory allocator. Taking
the lock required an allocation thus the thread acting within the exception
handler would get stuck.
* We save a few KiBs of memory per process since we don't have to keep the
thread name mapping around.
* Thread startup will be slightly faster since the first thing each new thread
did was to fill its own mapping.
Differential Revision: https://phabricator.services.mozilla.com/D139319
8a17aeccc2