[clang] Improve HeuristicResolver further so it can replace most of getApproximateType() in SemaCodeComplete (PR #156282)

[Nathan Ridge via cfe-commits]

HighCommander4 wrote:

The regression is caused by the additional of the following hunk to `resolveExprToType()`:

```c++
  if (const auto *CE = dyn_cast<CallExpr>(E)) {
    if (QualType Resolved = resolveTypeOfCallExpr(CE); !Resolved.isNull())
      return Resolved;
  }
```

If we have a `CallExpr`, but `resolveTypeOfCallExpr()` fails to resolve a type for it, we don't early-return but instead proceed to the rest of the function. Later in the function we call `resolveExprToDecls()`, which will call `resolveTypeOfCallExpr()` again (as part of calling `resolveTypeOfCallExpr()` + `resolveTypeToTagDecl()`). So we end up calling `resolveTypeOfCallExpr()` twice even though we know if won't produce a useful result.

As a result of this, a call to `resolveExprToDecls()` on an expression involved chained function calls will do an amount of processing that's expontential in the length of the chain.

Fix coming up.

https://github.com/llvm/llvm-project/pull/156282