We're only holding a read lock on the director, but we're updating
the nxt member concurrently. This should be acceptable as a performance
tradeoff - the only consequence is that round-robin is not strictly
going around - it may occasionally skip a backend or hand out the same
multiple times in a row.

the race is:

	thread	code

	A:	rr->nxt %= rr->vd->n_backend;
	// rr->nxt == rr->vd->n_backend - 1
	B:	rr->nxt++;
	// rr->nxt == rr->vd->n_backend
	A:	be = rr->vd->backend[nxt];
	// BOOM

should fix #2024