feat(prefer-set-size): add prefer-set-size rule

[bartlomieju]

This ports the prefer-set-size rule from oxlint's unicorn plugin as a native
Rust rule.

The rule flags a .length member access on an array produced by converting a
Set to an array, and recommends using Set#size instead. It catches both
spread conversions ([...set].length) and Array.from(set).length, where the
Set is either a direct new Set(...) or a const-declared variable
initialized to new Set(...).

It deliberately does not flag: optional chaining (?.length), computed access
([length] / ["length"]), non-length properties, new NotSet(...), bare
Set(...) without new, let/var-declared or destructured set variables,
Array.from(set, mapFn) (a second map-fn argument), Array?.from /
Array.from?.(), and NotArray.from.

The pass/fail fixtures are ported verbatim from the oxc reference rule:
https://github.com/oxc-project/oxc/blob/main/crates/oxc_linter/src/rules/unicorn/prefer_set_size.rs
(the autofix-only fixtures are dropped since deno_lint has no autofixer here).

Tagged RECOMMENDED (on by default) — flagging for maintainer sign-off.

[bartlomieju]

Verdict: LGTM — correct and well-tested. Recommend merge after the RECOMMENDED sign-off you flagged + rebase. Not auto-merging (new rule).

Careful two-pass port (collect const Set vars → check .length). I verified the tricky parts:

Correctness — verified

• Soundness via const-only: only const x = new Set(...) with a simple Ident binding is collected, so the variable can't be reassigned to a non-Set. let/var/destructured bindings are (correctly) not flagged. Right call.
• Hygiene: set_vars is keyed by Id (Atom + SyntaxContext), so a const set in one scope won't false-match a same-named binding in another. Good.
• Exclusions all hold: ?.length (parent is OptChainExpr), computed [length]/["length"], non-length props, new NotSet(), bare Set() w/o new, Array.from(set, mapFn) (is_array_from_call requires args.len() == 1), Array?.from/Array.from?.() (not a plain Expr::Member/Expr::Call), and multi-element arrays ([...set, x].length) — each is correctly rejected and covered by valid tests.
• get_set_node requires a single spread element for the array form ([set].length with a non-spread element is not flagged), and unwrap_parens is applied at both the object and set-node levels, so [...(set)].length still resolves.

Perf: two full traversals (collector + handler), each O(nodes), with O(1) HashSet lookups — fine for a per-file pass.

Housekeeping: RECOMMENDED sign-off (you flagged it); mergeable: UNKNOWN → rebase; docs .md not in diff.

No code changes requested — the const-only soundness and the exclusion set are exactly right. Just tagging + rebase.