Find path and route to next Trampoline node

In this commit, we expand our forwarding logic to do ad-hoc pathfinding
to subsequent Trampoline nodes, covering both blinded and unblinded
scenarios. We allow the forwards to use MPP, but we do not yet implement
inbound Trampoline MPP handling, nor retry logic.

We further modify our error propagation logic to handle Trampoline
forward HTLC failures eagerly, i.e. when we receive a failure from the
downstream node, we immediately fail or inbound MPP components.

Author: arik-so