linter: Rework only_used_in_recursion detection logic

[no-yan]

What version of Oxlint are you using?

No response

What command did you run?

No response

What does your .oxlint.json config file look like?

No response

What happened?

To improve detection, I propose modifying the algorithm to support a wider range of patterns, including multi-node expressions and cases with more complex references.

The current detection logic is conservative and based on the following conditions:

1. All references are used in the appropriate position (property or argument) of the recursive function call.
2. The call is a single-node invocation, e.g., recurfunc(arg), and not a multi-node expression, e.g., recurfunc(arg + 1).

Due to this, the following cases are not detected by the current implementation:

function recursive(arg) {
    return recursive(arg + 1);
}

function recursive(arg) {
    console.log(arg);
    return recursive(arg);
}

Clippy's implementation avoids issuing warnings for expressions with side effects.
We might consider adopting a similar approach in our implementation.

[Boshen]

@no-yan would like to work on this, or should I change this to good first issue?

[no-yan]

@Boshen Unfortunately, I didn't have the bandwidth to work on this for now.
Could you please unassign me?