Use an invalid table index as the callback for one-way functions

[nomeata]

this fixes #2938, and allows developers to write canisters that are
zero-downtime upgradeable, as long as they stick to making one-way calls
onlys (i.e. plain actor model)

[github-actions]

Comparing from 27476cb to 90705e4:
In terms of gas, no changes are observed in 3 tests.
In terms of size, no changes are observed in 3 tests.

[crusso]

Very nice but we should double check this doesn't somehow interfere with the work that @ggreif did with cleanup of the closure table after a trapping callback. Probably fine, but best to double check.

[nomeata]

Double checking is appreciated. I believe that the callback table wasn't used at all for one-ways, and no cleanup callback needed or used. (This also means that the “is it safe to upgrade” check was incomplete, as it would ignore outstanding callbacks to one-way calls.)

[crusso]

Looks ok to me, but I don't quite see why the error message has changed in the test? @ggreif have you taken a look at this?

[nomeata]

Because with this change, when the callback comes back, the system now complains about the callback is to be invalid. that's expected (and drun is even silent about this)

[ggreif]

Sorry for the delay, I am a bit behind. I'll have a look soon.

[nomeata]

Ping, @ggreif?

[ggreif]

This got buried again :-(, sorry!

I have double-checked, #2663 is compatible with this change, as it will try to access the continuation table in the same way as the other callbacks.

[ggreif]

LGTM!