Maybe make relay_digest_matches() not use tor_malloc()

The relay_digest_matches() is used in relay_crypt() and it is called for a huge portion of cells that come through the relay.

Roughly speaking, if a relay is at 10MB/s and with cells of size 514 bytes, we are talking about a bit less than 20k cells per second meaning more than a million tor_malloc(20) per minute. This is the place:

  backup_digest = crypto_digest_dup(digest);

I think we should find a way to avoid this especially in the fast path of tor in order to avoid memory fragmentation as much as possible.

We could either use the stack, a static value or memarea subsystem.