When using ChannelManager::pay_for_offer, we send an InvoiceRequest for each path in the Offer in parallel, paying the first Bolt12Invoice received in response and ignoring the rest. This is necessary because onion messages can be unreliable; we can't be sure a message was actually delivered. While this adds a degree of robustness to the offers flow, it is not enough. For instance, an Offer may only have one path so only one attempt would be made. Or ephemeral networking issues could result in all messages lost, etc.

Currently, OutboundPayments will fail any request that hasn't received a response after some StaleExpiration criteria. Before this is reached, we should attempt to retry the request as necessary. This should work across node restarts.