Fix override-features implementation

[pm47]

This is an alternative to #1498.

We were calling nodeParams.features from inside the channel, which is
problematic because we may have overriden those features for specific
peers. This is now fixed.

IncomingPacket.decrypt needs features to tell whether we are using
variable onion or not. What features should we use? Do we have an
unwritten rule that node-level features shouldn't be overriden? This
happens in the relayer and in the post-restart-htlc-cleaner.

[pm47]

Note that if we change the var_onion_optin after a restart, then the post-restart-htlc-cleaner task may not work. We could make PostRestartHtlcCleaner.checkBrokenHtlcs use channels.commitments.localParams.features to decrypt htlcs, wdyt?

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (master@56d216b). Click here to learn what that means.
The diff coverage is 100.00%.

@@            Coverage Diff            @@
##             master    #1549   +/-   ##
=========================================
  Coverage          ?   87.21%           
=========================================
  Files             ?      139           
  Lines             ?    10902           
  Branches          ?      428           
=========================================
  Hits              ?     9508           
  Misses            ?     1394           
  Partials          ?        0           

Impacted Files	Coverage Δ
...re/src/main/scala/fr/acinq/eclair/NodeParams.scala	89.65% <100.00%> (ø)
...c/main/scala/fr/acinq/eclair/channel/Channel.scala	85.95% <100.00%> (ø)
...c/main/scala/fr/acinq/eclair/channel/Helpers.scala	96.21% <100.00%> (ø)
...-core/src/main/scala/fr/acinq/eclair/io/Peer.scala	88.77% <100.00%> (ø)
...rc/main/scala/fr/acinq/eclair/io/Switchboard.scala	80.64% <100.00%> (ø)

[t-bast]

Do we have an unwritten rule that node-level features shouldn't be overriden?

I think so, yes. It feels to me that turning var_onion_optin off is asking for trouble...
There are some features that we may want to make mandatory at some point, and disallow users turning them off, don't you think? It will also allow us to simplify the code by removing some unnecessary if statements. I'm thinking that var_onion_optin is exactly a good candidate for that, there's too much risk of footgun for users if we let them turn it off (it's been battle-tested with 1.5 years in the wild and 0 drawbacks)