Ensure correct overload for diagnose() is called even in Transform context

[asl]

This ensures we do not see duplicate diagnostics from there.

Some small cleanup while there.

[asl]

Found this while working on #4829, full explanation is in #4829 (comment)

Long story short: when calling error() from inside Transformer we passed non-const pointer to Node and due to this wrong (SourceInfo-less) overload was selected.

[ChrisDodd]

It seems like from the stdout changes here, we're losing some non-duplicate error messages?

[asl]

It seems like from the stdout changes here, we're losing some non-duplicate error messages?

Not quite. We just visit the same (invalid) code twice despite error already being reported. As a result, multiple errors for the same source line were emitted.

For example:

type-in-expr.p4(5): [--Werror=type-error] error: 4 < H: structured annotation must be compile-time constant values
@foo[bar=4<H]
         ^^^
type-in-expr.p4(5): [--Werror=type-error] error: H: Type cannot be used here, expecting an expression.
@foo[bar=4<H]
           ^
type-in-expr.p4(6): [--Werror=type-error] error: Error while analyzing control p
control p<H>(in H hdrs, out bool flag)
        ^

So, here we are emitting two errors for the same line 5. And the diagnostics on line 6 is just an additional code inside learn() that checks error counts before and after and emit additional error if they are not same. In the majority of cases the latter is superflous.

After the patch the intended functionality is restored, only single error per source line is reported.