Cap concurrent password checks and count attempts before checking

Every argon2 run takes 64 MiB and nothing limited how many ran at once,
so parallel sign-in attempts could run the server out of memory (8 at
once used about 600 MB). At most two now run at once; at most 16
sign-ins wait for one, more get HTTP 429. 30 parallel sign-ins peaked
at 275 MB.

A sign-in attempt now counts toward the lockout before its password is
checked, so parallel attempts cannot get past it; a right password
takes its own attempt back. IPv6 addresses are locked out by /64.

api.go

@@ -232,7 +232,7 @@ func (a *App) login(w http.ResponseWriter, r *http.Request) {
 	if err != nil {
 		slog.Warn("login failed", "user", in.Username, "remote", ip, "reason", err.Error())
 		code := http.StatusUnauthorized
-		if errors.Is(err, errLocked) {
+		if errors.Is(err, errLocked) || errors.Is(err, errBusy) {
 			code = http.StatusTooManyRequests
 		}
 		writeJSON(w, code, map[string]string{"error": err.Error()})