Rather than using pthread_get/set_specific, just use gcc's __thread
annotation, which is noticeably faster and makes the code more obvious.

Also, just have one simplified struct called vgrnd, instead of trying to
split things up semantically. Those divisions were useful when this code
was split across several commit *messages*, but doesn't make as much
sense within a single file. This should make the code more clear and
provide a better example for implementers.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>