The code here would need some change anyway given planned change in
SPI_modifytuple semantics, since this executes after we've exited the
SPI environment.  But really it's better to just use heap_modify_tuple.

While at it, normalize use of SPI_fnumber: make error messages distinguish
no-such-column from can't-set-system-column, and remove test for deleted
column which is going to migrate into SPI_fnumber.  The lack of a check
for system column names is actually a pre-existing bug here, and might
even qualify as a security bug except that we don't have any trusted
version of plpython.