I understand the network like this.
1) Sender is searching the best route through all nodes to reach the target.
2) Sender is encrypting the selected route node by node (onion routing)
3) Sender send coins to the first node using a payment channel between sender and first node
4) The next node decrypts the follow up node and continue sending coins to the next node over the next payment channel
5) continue with 4 until the target node is reached.
In my understanding, whenever I send something over the payment channel, this transaction is completed and not reversible (this is maybe the point where I'm wrong).
So, when I have 10 nodes in my selected route, and node 5 in my list is down. How does the network make sure, that the 4 nodes before are doing a rollback to give back the coins to the sender, so the sender can try a different route.
That's not how payment routing actually works. That's just the messaging protocol to transmit some information between nodes.
The actual payment part works through Hashed Time Locked Contracts (HTLCs). So node A offers a HTLC output to node B. In order for node B to spend from that HTLC (thus being paid by node A), he needs to provide the preimage to some hash in addition to his own signature. After some timeout, node A can take the money back with a signature of his own.
In LN, the HTLCs are used to make the payments between the nodes. They use the same hash so they all have the same hash pre-image required. So if any one HTLC is spent, all HTLCs in the route can be spent since the hash pre-image is revealed during the spend. If the hash preimage is never revealed, then the coins are never sent to the other party. This makes it atomic.
As a shortcut (to avoid on chain transactions for each HTLC) LN uses the onion routing method you described. The hash preimage is given to the recipient by the sender and that is transmitted to each node in the route so that they all have the hash preimage. Once both nodes in the channel have the hash preimage, they resolve the HTLC off chain. If some nodes are not able to transmit the hash preimage to settle the HTLC off chain, then the node that was able to settle with one peer but not the other can settle on-chain, and in doing so, reveals the hash preimage for the remaining nodes to settle (either on chain or off chain, as necessary). That node is incentivized to settle on chain because he would actually be losing money by not settling.