…t funding

A user can be left with a channel wedged in a state that provably will
never resolve, with no transition to a sensible end state. The sequence:

How it gets stuck: BOLT 2 permits shutdown before channel_ready, so both
sides can finish a mutual close and store a signed close tx while still in
CHANNELD_AWAITING_LOCKIN; the channel reaches CLOSINGD_COMPLETE. That close
tx spends the funding output, so it only becomes usable once the funding tx
confirms. If the funding tx never confirms, the channel stays in
CLOSINGD_COMPLETE.

Why the state is held: while the funding inputs are unspent the
funding tx can still confirm (e.g. on rebroadcast), and if it does the
stored close resolves the channel normally, so the channel must be
kept until the funding is known dead.

After the double-spend is buried it is safe to clear: when a funding
input is spent by another tx and that spend is buried deep enough that a
reorg will not reverse it, the funding tx can never confirm. The funding
output will never exist, the channel can never resolve on chain, and it can
move to a clean end state.

This adds an xfail-strict reproducer that drives the channel into
CLOSINGD_COMPLETE behind a never-confirming funding tx, double-spends a
funding input, and asserts both edges: the channel must stay in
CLOSINGD_COMPLETE while the double-spend is shallow, and must reach a
resolved state once the spend is reorg-safe. The marker should be removed
once a fix exists.

Reproduces ElementsProject#9102.

Changelog-None