[Codegen] Self imported code compiled via legacy results in EVMC_STACK_OVERFLOW whereas via-ir results in a panic (underflow)

```
==== Source: s1.sol ====
import {f as g, g as h} from "s1.sol";
function f() pure returns (uint) { return 100 + h() - g(); }
==== Source: s2.sol ====
import {f as h} from "s1.sol";
function f() pure returns (uint) { return 2; }
function g() pure returns (uint) { return 4; }
==== Source: s3.sol ====
import "s2.sol";
contract C {
function foo() public pure returns (uint) {
return f() - g() - h();
}
}
// ----
// foo() -> FAILURE
```

To reproduce, copy the test file in `test/libsolidity/semanticTests/test.sol` and run

```
$ isoltest --show-messages semanticTests/test
```

Copy the reproducer into test/libsolidity/semanticTests/test.sol and run `isoltest --show-messages semanticTests/test`. Compare the legacy and via-ir code-generation paths for the self-imported functions, then verify the regression no longer produces EVMC_STACK_OVERFLOW or a panic from underflow in either mode.

Written by the indexing model from the issue text.