channeld: allow transient negative balance.

[rustyrussell]

Travis randomly picked up an error in test_feerate_stress:
BROKEN 0266e4598d1d3c415f572a8488830b60f7e744ed9235eb0b1ba93283b315c03518-channeld-chan#1: Cannot add htlc #0 10000msat to LOCAL (version a2541b9-modded)

This is because it hit an unlikely corner case involving applying multiple HTLCs
(similar to the previous c96cee9).

In this case, the test sends a 500,000,000 "balancing" setup payment L1->L2.
It waits for L2 to get the preimage (which is the when pay() helper returns),
but crucially, it starts spamming with HTLCs before that HTLC is completely
removed.

From L2's point of view, the setup HTLC is in state RCVD_REMOVE_REVOCATION;
gone from L1's commitment tx, but still waiting for the commitment_signed
from L1 to remove it from L2's.

Note that each side keeps a local and remove view of both sides' current
balances: at this point, L2's view is REMOTE: "500,000,000 to L1, 499,900,000
to L2", LOCAL: "500,000,000 to L1, 0 to L2".

L2 sends a 10,000 msat HTLC to L1: legal, since L1 will allow it,
then the commitment_signed. L1 sends the revoke-and-ack for this,
then belatedly follows with the commitment_signed which both completes the
removal of the setup HTLC and adds the new one.

But L2 processes the HTLCs in hashtable (i.e. random) order: so if it
tries to apply its own HTLC first, it freaks out because it doesn't have
funds in its local view.

Signed-off-by: Rusty Russell [email protected]
Changelog-Fixed: Unlikely corner case is simultanous HTLCs near balance limits fixed.

[cdecker]

ACK 90a9bd2