program: reallocate stake account on zero lamports

[2501babe]

instead of zeroing account data, we should realloc accounts to 0 size, to ease headaches of downstream programs dealing with stake accounts across multiple instructions in the same transaction

closes #71

[2501babe]

this might also be a good time to think about how we are going to do fuzzing, because it is a small simple change that slihgtly changes behavior. ideally we have some kind of runner via mollusk that tells us something changed between the two versions, we can say "ok this is fine" and adjust the things we expect to be different, and confirm nothing unexpected changed by accident. i have no idea yet what impling this kind of flow would be like since it depends on multiple versions of the program and idk if it can really be part of ci... idk think about it

[felix-asym]

I believe this change would break the stake cache logic in Agave.
See the comment here and the discussion here

The code for updating the stake cache currently assumes that the staking program doesn't change the owner of an account when it's lamports are zeroed out.

[joncinque]

Looks great to me! I took another looks through the stakes cache, and I believe this will be safe.

The stakes cache only stores stake accounts with delegations, and if lamports go to 0 OR if the account can no longer be deserialized as a stake account, it gets removed from the cache. So essentially we'll go from deserializing StakeState::Uninitialized and then failing to get its delegation.

We should be aware that this does create a DOS opportunity on stake accounts. If someone closes a stake account and then tops it up with lamports in the same transaction, it will be impossible to ever reuse it. We could relax the logic in withdraw to treat a zero-length account the same as a StakeState::Uninitialized -- what do you think?

[2501babe]

im gonna leave this in draft until im back, i havent thought deeply about downstream effects the zero-len accounts or the ability to withdraw from them. at least we need to add test cases to make sure some of the other instructions refuse to interact with them. its probably fine since i believe everything goes through get_stake_state() but yea

[joncinque]

The new commit looks great to me! And yeah definitely good to think about all of the other interactions