fix: DeliverMessageSend discards FSM read errors and issues conflicting Set+Delete for drained self-transfers

[amathxbt]

Bug 1 — swallowed FSM error

response, err := c.plugin.StateRead(c, &PluginStateReadRequest{...})
if err != nil {
    return &PluginDeliverResponse{Error: err}
}
// ensure no error fsm error
if response.Error != nil {
    return &PluginDeliverResponse{Error: err} // <-- err is nil here
}

err was already confirmed nil a few lines above, so when response.Error != nil this branch always returns Error: nil — the real FSM-level error is discarded. Execution then falls through to Unmarshal(fromBytes, from) etc. with empty byte slices (since response.Results is empty when response.Error is set), silently treating a failed read as an empty/zero-value account rather than surfacing what actually went wrong.

CheckTx handles the identical situation correctly a few lines above in the same file:

if err == nil {
    err = resp.Error
}

DeliverMessageSend should do the same. Fixed to return response.Error.

Bug 2 — conflicting Set+Delete on the same key

if from.Amount == 0 {
    resp, err = c.plugin.StateWrite(c, &PluginStateWriteRequest{
        Sets: []*PluginSetOp{
            {Key: feePoolKey, Value: feePoolBytes},
            {Key: toKey, Value: toBytes},
        },
        Deletes: []*PluginDeleteOp{{Key: fromKey}},
    })
}

For a self-transfer (msg.FromAddress == msg.ToAddress) that drains the balance to exactly zero, fromKey == toKey (the code even special-cases this above with to = from). This write batch therefore contains both a Set and a Delete for the identical key, which is apply-order-dependent in the underlying store. On a blockchain this is a determinism risk: if different store implementations (or even the same implementation under different internal orderings) resolve Set-vs-Delete-on-same-key differently, nodes could diverge on whether the account ends up deleted or persisted with a zero balance.

Fixed to only include the to Set when it targets an account distinct from from, so a drained self-transfer results in a single unambiguous Delete.

Testing

Traced both code paths by hand against the existing (correct) CheckTx pattern and the self-transfer branch above it. No existing test suite in this repo to extend.