Make the existence test depend on the nodelist, not on node values

[ArsalLivoroi]

Refs #91. This covers part of it, see the note at the end.

RFC 9535 section 2.3.5.2 makes a test expression true when the nodelist it
produces is non-empty. The value of the nodes does not take part in it.

src/query/atom.rs looked inside the value instead:

let struct_check = |s: &T| {
    if let Some(arr) = s.as_array() { !arr.is_empty() }
    else if let Some(obj) = s.as_object() { !obj.is_empty() }
    else if let Some(str) = s.as_str() { !str.is_empty() }
    else { true }
};

So $[?@.x] dropped objects whose x is an empty array, an empty object or an
empty string, while keeping null, 0 and false:

{"x": [1,2]}   kept
{"x": []}      dropped, but x exists
{"x": {}}      dropped, but x exists
{"x": ""}      dropped, but x exists
{"x": null}    kept
{"x": 0}       kept
{"x": false}   kept

There was a second point in the same expression: Data::Refs(elems) => elems .iter().map(|v| v.inner).all(struct_check) used all, so a single empty node
made the whole test false even when the nodelist had several members.

Both are replaced by a nodelist-emptiness check.

What this does not fix

filter, nested stays red, so I have not marked this as closing #91. The
remaining problem looks separate, and specific to a filter selector applied
directly to @:

$[?@[0]]        works
$[?@[*]]        works
$[?@.v[?@>1]]   works
$[?@[?@>1]]     returns nothing

I have not looked into that one yet.

One question: was struct_check there for a reason I am not seeing? Every unit
test and the compliance suite pass without it, but you would know better whether
something depends on the old behaviour.

[besok]

thank you for the pr. i missed this part apparently in the spec.