Fold in leios testnet bugfixes from 20260708 stuck sync incident

[nfrisby]

This PR fixes three bugs.

• We were accidentally using the default method headerContainsLeiosCert for the HFC, so we essentially never interepted MsgRollForward as an offer. I have no idea how the manual testing had done for that original PR didn't fail despite that.
• The pruning logic for the set of EBs acquired was incorrect: it was assuming that all the announcements were necessarily in the VolDB. But---somewhat counterintuitively---our syncing node is downloading the EB before it even downloads the announcer's RB. So this pruning logic was instantly deleting the EB.
• There was a deeply fundamental one-line bug in the implementation of the pipelined LeiosFetch client: if some unprocessed responses had arrived, we would sometimes (at the very last possible moment) discard a request instead of actually sending it to the upstream peer. But the LeiosFetch decision logic's accounting wasn't informed, so it eventually concluded too many requests were outstanding (even though they were never actually sent) for it to be able to send any more.

This code has been running on the testnet for 12 hrs now, without exhibiting the stuck syncs (which has been perfectly reliable prior to these).