Fix revert account destroyed with nonexistent target

[LindaGuiga]

We noticed that the create2collisionSelfdestructedRevert tests from the evm-tests test suite were failing with a Kernel panic. This is because in revert_account_destroyed, we access the target's balance to undo the balance change carried out in SELF_DESTRUCT. But in the tests, the target account had to be created when carrying out a SELF_DESTRUCT, and the revert therefore already deleted it. In that case, mpt_read_state_trie returns 0 (corresponding to a nonexistent account), and we should not try to change the target's balance.

This PR therefore ensures that when a target account is nonexistent, we do not try to change its balance.

[wborgeaud]

Good catch!
Your fix should work, but I think the underlying issue here is that in sys_selfdestruct we do %journal_add_account_destroyed before doing %journal_add_account_created (in %add_eth).
It should be the other way around and looking at the revm implementation, that's what they do too.
I've pushed this change here https://github.com/0xPolygonZero/plonky2/tree/wb/testing_selfdestruct and it makes the create2collisionSelfdestructedRevert tests pass, although I haven't run the full suite with it yet.
What do you think?

[LindaGuiga]

@wborgeaud Thank you! Your solution does seem better! I ran it on 6 tests that were failing due to this issue, and they all pass with your solution as well (all the create2collisionSelfdestructedRevert and create2collisionSelfdestructedOOG). I also haven't run the full suite with it, but it looks good to me

[wborgeaud]

@LindaGuiga Great! Just opened #1287 with the fix.