We previously relied on `context.child` to check whether we already had a
relay handler for a given payment_hash.

Unfortunately this could return an actor that is currently stopping itself.
When that happens our relay command can end up in the dead letters and the
payment will not be relayed, nor be failed upstream.

We fix that by maintaining the list of current relay handlers in the
NodeRelayer and removing them from the list before stopping them.
This is similar to what's done in the MultiPartPaymentFSM.